MARC2 Puzzle Inspector

▶ Task 6: "A sail catching the wind snaps taut against its mast" k=1 alts: 13/15 025d127b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input grid contains one or more outlined parallelogram-like shapes made of a single non-black color each, set against a black background. Each shape is an open outline (not filled) forming a slanted quadrilateral: the top and bottom edges are horizontal line segments, while the left and right edges slant diagonally downward to the right. The right edge and left edge have the same diagonal slope. The bottom-right corner of each shape is where the right edge meets the bottom edge.

Do: For each colored shape independently: (1) Identify the bottommost row of the shape and the rightmost column occupied by any pixel of the shape. (2) Every pixel that is NOT on the bottommost row of the shape AND NOT on the rightmost column of the shape gets shifted one cell to the right. (3) Pixels on the bottommost row or on the rightmost column remain in place. This has the effect of straightening the right edge of the parallelogram — the diagonal right side becomes more vertical at the bottom while the overall shape narrows its slant by one column.

Grid: The grid dimensions remain unchanged from input to output. The background is black (.). Shapes do not touch the grid edges and do not overlap each other. Each shape occupies a contiguous rectangular bounding region. Colors 0-9 are possible but only non-black colors form the shapes.

Figurative Descriptions (16 variants)

"A sail catching the wind snaps taut against its mast"

120b: k=1

See: Against a dark sea, colored sails billow in the wind — each one a slanted, hollow shape whose fabric stretches diagonally from its mast line on the right. The sails are slack, their left edges pulling away at an angle, making them lean like parallelograms rather than standing upright.

Do: A gust of wind blows from the left, pressing each sail flat against its mast. The mast (the rightmost edge) and the boom (the bottom edge) are rigid — they hold firm. But every loose piece of canvas not anchored to the mast or boom gets pushed one step rightward toward the mast, tightening the billow. The sail becomes tauter, its silhouette a little more upright, as the slack fabric snugs closer to the rigid frame.

Grid: The dark ocean fills the grid unchanged. Each sail is a single color's hollow outline floating in its own patch of sea. Sails never touch the edges of the world or overlap one another. The mast and boom hold fast while the free canvas shifts windward.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (8136 chars)

We need to infer rule. Input shows a shape: a hollow outline of a color (A in example) shaped like a slanted parallelogram? Let's visualize example 1 input grid 8x9? Actually given 8 rows, 9 columns? Let's count: first line ". . . . . . . . ." =9 columns. rows 8. The shape: azure A forms a diagonal line from left to right? Let's map coordinates (row, col). Row1 index0? Row0 all dots. Row1: ". A A A A A . . ." columns 1-5 A. Row2: ". A . . . . A . ." columns1 and6 A. Row3: ". . A . . . . A ." columns2 and7 A. Row4: ". . . A . . . . A" columns3 and8 A. Row5: ". . . . A A A A A" columns4-8 A. So shape is like a hollow right-leaning parallelogram with right edge vertical? Actually mast is rightmost edge (vertical line) and bottom edge (boom) horizontal? In input, the rightmost column of shape is at column8 row4 and row5? Row5 has A at col8 (rightmost). Row4 also at col8. So right edge vertical from row4 to row5? Actually row4 col8 A, row5 col8 A. That's vertical line of length2. Bottom edge is row5 columns4-8 A (horizontal). So shape is a hollow outline of a parallelogram leaning leftwards? The left edge is diagonal from (row1,col1) down-right to (row4,col3). That's a diagonal line.

Now output: shape shifted? Output rows: Row1: ". . A A A A A . ." columns2-6 A (shifted right by 1?). Input row1 had A at cols1-5. Output row1 has at 2-6. So moved right one. Row2 unchanged? Input row2 had A at col1 and6. Output row2 same? Output row2: ". . A . . . . A ." columns2 and7? Wait output row2: ". . A . . . . A ." that's col2 and7. Input row2 had col1 and6. So shifted right by1. Row3 input col2 and7; output col3 and8? Output row3: ". . . A . . . . A" columns3 and8. Yes shift right 1. Row4 input col3 and8; output col4 and8? Output row4: ". . . . A . . . A" columns4 and8. Left edge moved right, but right edge stayed at col8 (mast). Row5 input bottom row col4-8; output same col4-8 (unchanged). So rule: shift all non-mast, non-bottom edge cells one step right, but keep mast (rightmost vertical edge) and bottom edge fixed.

Essentially "press sails flat": move all canvas cells that are not on the rightmost column of shape (mast) nor on the bottom row of shape (boom) one step right.

Thus for test input with Y shape similar but larger? Let's map.

Test input grid 10 rows, 9 columns? Actually first line ". . . . . . . . . ." 9 columns. rows 10. Let's locate Y shape.

Row0 all dots.
Row1: ". Y Y Y Y Y Y . . ." columns1-6 Y.
Row2: ". Y . . . . . Y . ." columns1 and7 Y.
Row3: ". . Y . . . . . Y ." columns2 and8 Y.
Row4: ". . . Y . . . . . Y" columns3 and9? Wait only 9 columns, index0-8. Row4 string: ". . . Y . . . . . Y" that's 10 entries? Let's count: tokens: '.' '.' '.' 'Y' '.' '.' '.' '.' '.' 'Y' =10 columns. But earlier rows have 9 columns. Something off. Let's recount each row tokens.

Given test input lines:

". . . . . . . . . ."
". Y Y Y Y Y Y . . ."
". Y . . . . . Y . ."
". . Y . . . . . Y ."
". . . Y . . . . . Y"
". . . . Y Y Y Y Y Y"
". . . . . . . . . ."
". .

... [5187 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.94)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 35: "Seeds dropped in soil grow into sedimentary rock layers, eac ... [44 chars truncated]" k=1 alts: 14/15 0f63c0b9

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input is a 15x15 grid with a black background containing several single colored pixels (dots) scattered at various positions. Each dot is a different color. The dots' row positions determine how the output grid is partitioned into horizontal bands.

Do: 1. Identify all colored dots in the input grid and record each dot's color and row index. Ignore column positions. 2. Sort the dots by row index from top to bottom. 3. Partition the 15 rows into consecutive bands by computing midpoints between adjacent dots' row indices. The boundary between two adjacent dots at rows r1 and r2 is floor((r1 + r2) / 2) + 1, meaning the first band ends at row floor((r1+r2)/2) and the next band starts at floor((r1+r2)/2)+1. The topmost band starts at row 0 and the bottommost band ends at row 14. 4. For each band, fill it with its dot's color as follows: a. The row where the dot originally sat becomes a full solid horizontal line (all 15 cells filled with the dot's color). b. The first and last rows of the entire grid (row 0 and row 14), if they fall within this band, also become full solid horizontal lines of the dot's color. c. All other rows in the band become border rows: only the leftmost cell (column 0) and rightmost cell (column 14) are filled with the dot's color; the interior cells (columns 1-13) remain black.

Grid: Both input and output grids are 15x15. The background color is black (.). The input contains exactly N single-colored pixels (N varies per puzzle instance) placed at distinct positions. The output uses the full 15x15 grid, partitioned into N horizontal bands with no gaps or overlaps, each rendered as a rectangular frame segment in the corresponding dot's color.

Figurative Descriptions (16 variants)

"Seeds dropped in soil grow into sedimentary rock layers, each mineral claiming its stratum of the earth."

120b: k=1

See: A column of dark earth stretches before you. Scattered at different depths, tiny mineral seeds sit buried — each a different color, each waiting to crystallize. Their depth is everything; their left-right position is forgotten the moment they begin to grow.

Do: Each seed awakens and claims a horizontal stratum of the earth proportional to the space between its neighbors above and below. Where the seed was planted, a solid vein of ore runs the full width of the earth — the densest concentration. At the very surface and the very bottom of the world, exposed strata also harden into solid bands. Everywhere else within a stratum, the mineral only lines the walls — crystallizing along the left and right edges while the interior remains hollow dark rock, like a geode split open.

Grid: The earth is a 15-layer-deep cross-section. Each mineral seed is a single colored speck in the dark. Once they crystallize, the entire cross-section is claimed — no dark void is left unowned. Each stratum is a hollow corridor: solid colored ceiling where the seed sat, colored walls on left and right edges, and darkness filling the interior between them.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (14101 chars)

We need to understand rule. From example: Input 15x15 grid with three colored seeds: R at row4 col7? Let's index rows 0-14. Input shows R at row3 (0-index) column6? Actually row4 line: ". . . . . . R . . . . . . . ." So column6 (0-index). O at row7 col8? line: ". . . . . . . . O . . . . . ." column8. A at row12 col3? line: ". . . A . . . . . . . . . . ." column3.

Output: shows horizontal bands. For each seed, a band spanning rows? Let's examine output rows.

Rows 0-14 output given.

Row0: all R across entire row. Row1: R at col0 and col14, rest .? Actually row1: "R . . . . . . . . . . . . . . R". So R at edges only. Row2 same. Row3: all R. Row4: R at edges only. Row5: R at edges only. Row6: O at edges only? Row6: "O . . . . . . . . . . . . . . O". Yes O edges. Row7: all O. Row8: O edges only. Row9: O edges only. Row10: A edges only? Row10: "A . . . . . . . . . . . . . . A". Row11 same. Row12: all A. Row13: A edges only. Row14: all A.

So pattern: For each seed, we create a horizontal stratum whose height equals distance between neighboring seeds (including top and bottom boundaries). The seed row becomes the "solid ceiling" (full row of that color). The bottom of stratum becomes solid floor? Actually bottom row of stratum also solid? In example, for R stratum: top boundary is row0 (top of world) solid R. Then rows1-2 are walls only (R at leftmost and rightmost). Then row3 (seed row) is solid R. Then rows4-5 are walls only. Then row6 is start of next stratum O with walls only? Wait O stratum top row is row6? Actually O stratum top row is row6 walls only, row7 solid O, row8-9 walls only, then row10 start of A stratum walls only, row12 solid A, etc. At bottom, final stratum bottom row (row14) solid A.

Thus each stratum consists of: top solid row (if at top boundary) else walls only? Let's examine O stratum: The seed O is at row7. The rows above O within its stratum are rows6 (walls only). The top of O stratum is row6 (walls only) not solid. The solid row appears at seed row (row7). The bottom of O stratum is row10? Actually bottom of O stratum is row9 walls only, then next stratum A starts at row10 walls only. So bottom solid row appears at bottom boundary (row14) for last stratum. For first stratum, top boundary solid row appears at row0. So rule: At world top and bottom, the stratum's outermost rows become solid across full width. For interior strata, only the seed row is solid; the rows at the boundaries between strata are walls only (colored at leftmost and rightmost columns). So each stratum height = distance between seeds (including boundaries). The seed row is somewhere within stratum: its position relative to top? In example, R seed at row3, stratum height from top (row0) to next seed O row7? Actually distance between R seed row3 and O seed row7 is 4 rows (rows4-7?). But stratum includes rows0-5? Wait R stratum includes rows0-5. That's 6 rows. Let's compute: top boundary row0, seed row3, next seed row7. The stratum extends from top boundar

... [11122 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 45: "Small tribes without enough members to survive are absorbed ... [20 chars truncated]" k=1 alts: 2/15 12eac192

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains black (.) background cells and non-black colored cells (which may include any combination of B, O, A, X, and possibly other colors). The non-black cells form connected components of same-colored cells using 4-connectivity (up, down, left, right — not diagonal). These components vary in size from single isolated cells to larger groups of 3 or more connected same-colored cells.

Do: 1. Identify every non-black cell in the grid and determine its color. 2. For each non-black cell, find the entire connected component it belongs to, where a connected component is defined as all cells of the same color reachable via 4-directional adjacency (up, down, left, right). 3. Count the number of cells in each connected component. 4. If a connected component has 3 or more cells, leave all its cells unchanged (they keep their original color). 5. If a connected component has fewer than 3 cells (i.e., size 1 or 2), replace every cell in that component with green (G). 6. All black (.) cells remain black.

Grid: The output grid has the same dimensions as the input grid. The background is black (.). The coordinate system is row-column with the origin at the top-left. There are no fixed size constraints — the grid can be any rectangular dimensions.

Figurative Descriptions (16 variants)

"Small tribes without enough members to survive are absorbed by the green forest."

120b: k=1

See: A dark landscape dotted with settlements of various colored tribes — blue villages, orange encampments, azure outposts, grey hermitages. Some tribes have gathered in strength, their people clustered tightly together along roads running north-south and east-west. Others are lonely — a single wanderer here, a pair of companions there, cut off from their kin.

Do: The green forest is always growing, always hungry. It respects strength in numbers — any tribe that has gathered at least three of its people together in an unbroken chain stands firm, their settlement walls holding back the encroaching woods. But the stragglers, the isolated outcasts and small pairs who couldn't find enough of their own kind to form a viable community, are swallowed by the forest. Their colors fade beneath the canopy, replaced by green. The darkness between settlements is too barren even for the forest — it remains untouched.

Grid: The landscape is a rectangular territory viewed from above, dark ground forming the empty wilderness between settlements. Tribe members touch only along the four cardinal roads — diagonal neighbors are strangers. The forest's green conquest replaces the lost, while the dark ground and the strong tribes remain exactly as they were.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10710 chars)

We need to infer rule: From example, we have input and output. Let's map.

Input grid (example1) size? Let's write rows:

Row1: B O O B . A . X
Row2: B O O B B . B .
Row3: A A . . O O O O
Row4: . B . . . . B B
Row5: X . A . B . B B

Output rows:

Row1: G O O B . G . G
Row2: G O O B B . G .
Row3: G G . . O O O O
Row4: . G . . . . B B
Row5: G . G . G . B B

We need to see which cells changed to G (green). Others unchanged.

Identify groups of same color (non-green) that are at least size 3 in orthogonal connectivity. Those groups stay same; others become G? Let's test.

Look at input: color B (blue). Where are B cells? Coordinates (row,col) 1-index.

Row1 col1 B, col4 B.
Row2 col1 B, col4 B, col5 B, col7 B.
Row4 col2 B, col7 B, col8 B.
Row5 col5 B, col7 B, col8 B.

Also B at row5 col? row5 col1 X not B.

Now connectivity: Let's see clusters.

Cluster1: B at (1,1) adjacent to (2,1) vertically => cluster includes (1,1),(2,1). Also (2,1) adjacent to (2,2) O not B. (2,1) adjacent to (3,1) A not B. So cluster size 2.

Cluster2: B at (1,4) adjacent to (2,4) B, (2,5) B adjacent to (2,4) horizontally, also (2,5) adjacent to (3,5) O not B. Also (2,4) adjacent to (2,5) and (3,4) . not B. So cluster includes (1,4),(2,4),(2,5) size 3.

Cluster3: B at (2,7) adjacent to (3,7) O not B, (2,8) . not B, (1,7) . not B, (3,7) O. So isolated size1.

Cluster4: B at (4,2) adjacent to (5,2) . not B, (4,1) . not B, (3,2) A not B, (4,3) . not B. size1.

Cluster5: B at (4,7) adjacent to (4,8) B, (5,7) B, (3,7) O, (4,6) . So cluster includes (4,7),(4,8),(5,7),(5,8) also (5,5) B? Wait (5,5) B adjacent to (5,7) not directly (col6 is .). So not connected. So cluster size 4.

Cluster6: B at (5,5) isolated? Adjacent (4,5) . , (5,4) . , (5,6) . , (6,5) none (grid only 5 rows). So size1.

Thus clusters of size >=3: cluster2 (size3) and cluster5 (size4). According to description, those stay same (remain B). Others become G (green). Let's see output: positions that were B and became G?

Check output row1 col1: G (was B) yes cluster1 size2 -> turned G. Row2 col1: G (was B) yes. Row2 col7: G (was B) yes cluster3 size1 -> G. Row4 col2: G (was B) yes cluster4 size1 -> G. Row5 col5: G (was B) yes cluster6 size1 -> G. Also row5 col1? was X unchanged? Output row5 col1 G? Wait output row5 col1 is G (input X). Actually input row5 col1 is X (grey). Output row5 col1 G. So not only B changed. Also row1 col8 X became G. Row2 col8 . unchanged? output row2 col8 . (same). Row3 col1 A became G? Input row3 col1 A, output row3 col1 G. Row3 col2 A became G. Row5 col3 A became G. So many non-B turned to G.

Thus rule not just about B clusters. Maybe any color that does NOT belong to a cluster of size >=3 becomes G, regardless of original color. Let's test: Colors present: O (orange), A (azure), X (grey), . (black). Check clusters for each color.

We need to find clusters of each color (excluding green). If a cluster size >=3, keep original colors; else turn to G.

Let's test O clus

... [7770 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.03)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 47: "Archaeologists piece together a shattered mosaic from fragme ... [36 chars truncated]" k=2 alts: 8/15 137eaa0f

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is an 11x11 grid with a black background. Scattered across the grid are several grey (X) cells and several colored (non-black, non-grey) cells. Each colored cell is spatially close to exactly one X cell. The X cells act as anchor points, and each has one or more colored cells nearby (within one step in row and/or column). Together, the colored cells and X cells encode fragments of a single 3x3 pattern.

Do: 1. Identify all X cells in the input grid. These are anchor points representing the center cell (row 1, column 1) of a shared 3x3 output grid. 2. Initialize a 3x3 output grid filled with black (.) and place X at position (1,1) — the center. 3. For each colored cell (any color other than black or grey), find the nearest X cell (by Manhattan distance or Euclidean distance — each colored cell will be unambiguously closest to exactly one X). 4. Compute the row and column offset of the colored cell relative to its nearest X: (delta_row, delta_col) = (colored_row - X_row, colored_col - X_col). 5. Map the offset to the 3x3 output grid by adding it to the center position: output position = (1 + delta_row, 1 + delta_col). 6. Write the colored cell's color into that output position. 7. Repeat for all colored cells. The result is the completed 3x3 output grid.

Grid: The input grid is 11x11 with a black (.) background. The output grid is always 3x3. The center cell of the output is always X (grey). Each offset from an anchor X falls within the range [-1, +1] in both row and column, so all fragments map into valid positions within the 3x3 output. No two colored cells map to the same output position.

Figurative Descriptions (16 variants)

"Archaeologists piece together a shattered mosaic from fragments found near buried survey stakes."

120b: k=2

See: Across a dark excavation site, several grey survey stakes have been driven into the earth. Around each stake, colored mosaic fragments lie half-buried in the soil — each shard resting just a step or two from its stake, as though a single mosaic tile was smashed and its pieces scattered to different corners of the dig site, each piece coming to rest near a different stake.

Do: The archaeologists realize every stake marks the same center point of the original mosaic — each stake is a ghost of the same heart-stone. The colored shards near each stake remember where they sat relative to that center. Gather every shard, note how it leans from its stake — above, below, left, right, diagonal — and lay it back into its rightful place around a single restored heart-stone. When every fragment has come home, the mosaic is whole again.

Grid: The dig site is a wide dark field. The restored mosaic is a small 3×3 tile with the grey heart-stone always at its center. Every shard nestles within one step of the center — no fragment was flung farther than arm's reach. No two shards compete for the same seat in the mosaic.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (8921 chars)

We need to infer rule. From description: stakes are grey (X) representing center of each 3x3 mosaic? Actually grey heart-stone always at center. In examples, output is 3x3 grid with center being X? Let's examine Example1.

Input grid 11x11. Output:
M R R
O X O
G G G

So output 3x3. Center is X (grey). So X is the stake (center). The other colors are placed relative to X based on their positions relative to any X in input? There may be multiple X's in input, each representing a stake (center). The colored shards near each stake are within one step (Manhattan? Chebyshev distance <=1). They are placed accordingly in output mosaic.

In Example1 input, locate X's. Let's map coordinates (row, col). I'll index rows 0-10, cols 0-10.

Rows:
0 all .
1 all .
2: ". M . . . . . . . . ." => M at (2,1)
3: ". . X . O X O . . . ." => X at (3,2), O at (3,4), X at (3,5), O at (3,6)
4 all .
5 all .
6 all .
7: ". . . . . . . X . . ." => X at (7,7)
8: ". . R R . . G G G . ." => R at (8,2),(8,3); G at (8,6),(8,7),(8,8)
9: ". . X . . . . . . . ." => X at (9,2)
10 all .

So there are multiple X's: at (3,2), (3,5), (7,7), (9,2). The description says each stake marks same center point of original mosaic — each stake is a ghost of same heart-stone. So we need to combine fragments from all stakes into one mosaic.

We need to collect colored fragments that are within one step of any X (including diagonals). For each such fragment, note its relative position to its X (dx,dy). Then place that color at same relative position around a single central X in output.

But there may be multiple X's with overlapping relative positions; but rule says no two shards compete for same seat, so they are consistent.

Let's test: For each X, look at neighboring cells (8 directions). Gather colors.

Take X at (3,2). Its neighbors:
(2,1)=M (up-left) => relative (-1,-1) => top-left of mosaic => output top-left should be M. Indeed output top-left is M.
(2,2)=. (up)
(2,3)=. (up-right)
(3,1)=. (left)
(3,3)=. (right)
(4,1)=. (down-left)
(4,2)=. (down)
(4,3)=. (down-right)
So only M.

X at (3,5). Neighbors:
(2,4)=. (up-left)
(2,5)=. (up)
(2,6)=. (up-right)
(3,4)=O (left) => relative (0,-1) => middle-left => output middle-left should be O. Indeed output middle-left is O.
(3,6)=O (right) => middle-right => output middle-right O. Yes.
(4,4)=. (down-left)
(4,5)=. (down)
(4,6)=. (down-right)

X at (7,7). Neighbors:
(6,6)=. (up-left)
(6,7)=. (up)
(6,8)=. (up-right)
(7,6)=. (left)
(7,8)=. (right)
(8,6)=G (down-left) => relative (+1,-1) => bottom-left => output bottom-left G. Indeed output bottom-left G.
(8,7)=G (down) => bottom middle => output bottom middle G. Yes.
(8,8)=G (down-right) => bottom-right G. Yes.

X at (9,2). Neighbors:
(8,1)=. (up-left)
(8,2)=R (up) => relative (-1,0) => top middle? Wait relative to X at (9,2): up is (8,2) => that's top middle of mosaic => output top middle should be R. Indeed output top middle is R.


... [6119 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.33)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 58: "A banner unfurls along the flagpole that holds it" k=3 alts: 7/15 15696249

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input is a 3×3 grid. Exactly one row or one column is filled with a single uniform color (all three cells are the same color). The remaining cells contain a mix of other colors.

Do: 1. Identify the uniform line: find the row or column where all three cells share the same color. Note whether it is a row (top=0, middle=1, bottom=2) or a column (left=0, middle=1, right=2). 2. Create a 9×9 output grid filled entirely with black (.). 3. If the uniform line is a row at position R (0, 1, or 2): copy the entire 3×3 input into the output three times side by side, placing them in output rows R×3 through R×3+2, spanning all 9 columns. 4. If the uniform line is a column at position C (0, 1, or 2): copy the entire 3×3 input into the output three times stacked vertically, placing them in output columns C×3 through C×3+2, spanning all 9 rows.

Grid: The input grid is always 3×3. The output grid is always 9×9 (a 3×3 arrangement of 3×3 blocks). The background fill is black (.). The 3×3 input is tiled three times along the axis indicated by the uniform row or column, placed in the corresponding strip of the 9×9 grid.

Figurative Descriptions (16 variants)

"A banner unfurls along the flagpole that holds it"

120b: k=3

See: A small 3×3 tapestry hangs in a dark hall. One thread—a single row or column—is woven from a single pure color, forming the flagpole that holds the whole design together. The rest of the weave is a patchwork of different colors.

Do: The flagpole knows which wall it belongs to. A row-pole is horizontal and knows its height; a column-pole is vertical and knows its side. The tapestry longs to be seen, so it unfurls itself three times along the direction of its pole, like a banner repeating in the wind. Wherever the banner doesn't reach, the dark hall remains—empty black stone stretching into shadow.

Grid: The dark hall is a 9×9 chamber of black stone. The flagpole's position decides which strip of the chamber gets decorated: a horizontal pole at the top, middle, or bottom places three copies side by side in the matching horizontal band; a vertical pole on the left, center, or right stacks three copies in the matching vertical band. The rest of the chamber stays dark and bare.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (6412 chars)

We need to understand the puzzle. Input is 3x3 grid. Output is 9x9 grid. The description: there is a "flagpole" which is a single row or column of uniform color (all same). In examples, we need to identify which row or column is uniform.

Example1 input:
B B B
M R R
R R M

Row0: B B B all same -> uniform row (horizontal pole) at top? Actually row0 is uniform B. So horizontal pole at top. Output: The top three rows (band) contain three copies side by side? Let's see output: first row: B B B B B B B B B (9 cells). That's three copies of row0 concatenated horizontally: B B B repeated three times. Indeed row0 repeated three times horizontally across the 9 columns. Row1 output: M R R M R R M R R (that's row1 repeated three times). Row2 output: R R M R R M R R M (row2 repeated three times). Then rows 3-8 are all '.' (black). So the flagpole is horizontal at top, so we fill the top band (first 3 rows) with the 3x3 pattern repeated horizontally three times (i.e., tile horizontally). The rest rows are black.

Example2 input:
R Y G
R G Y
R G Y

Check uniform column? Column0: R,R,R all same -> vertical pole on left. Indeed column0 uniform R. Output: The leftmost three columns (band) contain three copies stacked vertically? Let's see output: 9x9 grid. The first three columns contain the original 3x3 pattern repeated vertically three times. Let's examine output rows: Row0: R Y G . . . . . . (first three columns R Y G). Row1: R G Y . . . . . . (R G Y). Row2: R G Y . . . . . . (R G Y). Row3: R Y G . . . . . . (again). Row4: R G Y ... Row5: R G Y ... Row6: R Y G ... Row7: R G Y ... Row8: R G Y ... So indeed the 3x3 pattern is repeated vertically three times in the leftmost three columns. The rest columns are '.'.

Example3 input:
G B M
G M B
G B M

Check uniform column? Column0: G,G,G all same -> vertical pole left again. Output similar: leftmost three columns repeated vertically three times. Indeed output shows same pattern repeated vertically.

Thus rule: Identify the uniform row (all same color) or uniform column (all same). If uniform row, repeat the whole 3x3 pattern horizontally three times, placing in the rows corresponding to that row's position? Wait in example1, uniform row is top row (row0). Output placed repeated pattern in rows 0-2 (the top band). So the band corresponds to the row index of the uniform row: top, middle, bottom. Since uniform row is at top, we fill rows 0-2. If uniform row were middle (row1), we would fill rows 3-5? Actually "horizontal pole at the top, middle, or bottom places three copies side by side in the matching horizontal band". So if uniform row is middle (row1), the band would be rows 3-5? Wait the band is a horizontal band of height 3. The 9x9 grid is divided into three horizontal bands: top rows 0-2, middle rows 3-5, bottom rows 6-8. The pole's position (top/middle/bottom) determines which band gets the repeated pattern. For a uniform row at top, we fill top band. For uniform row at middle, fill middle band. For 

... [3454 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.56)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 85: "A snake slithers through a window frame, its body swaying si ... [11 chars truncated]" k=4 alts: 7/15 1c56ad9f

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid has a black background with a single colored rectangle drawn in one non-black color. The rectangle is defined by solid border rows at the top and bottom, and solid border columns at the left and right, forming a grid-like structure. The interior is divided into cells by internal horizontal and vertical lines of the same color, with black cells filling the spaces between the lines. The rectangle can vary in size, position, number of internal divisions, and color.

Do: Identify the rectangular structure and its top and bottom border rows. Let the bottom border row have distance d=0. For each row of the rectangle, compute d = (bottom_row - current_row). Shift the entire row's colored content horizontally according to d mod 4: if d mod 4 is 0, no shift; if d mod 4 is 1, shift left by 1; if d mod 4 is 2, no shift; if d mod 4 is 3, shift right by 1. Shifting means moving every cell in that row (both colored and the black cells between colored cells) by 1 position in the indicated direction. Vacated positions become black and content that moves fills its new position. All rows outside the rectangle remain unchanged.

Grid: The grid dimensions match the input exactly. The background is black (.). The colored rectangle occupies a contiguous region and is always axis-aligned. The grid is large enough that horizontal shifts of ±1 do not push content outside the grid boundaries.

Figurative Descriptions (16 variants)

"A snake slithers through a window frame, its body swaying side to side."

120b: k=4

See: Against the darkness, a colored window frame floats — a lattice of panes divided by mullions, like a church window suspended in the void. The frame is rigid and geometric, every crossbar perfectly aligned.

Do: A serpentine wave passes through the window from bottom to top. The bottom edge of the frame is anchored, but as the wave ripples upward, each horizontal bar of the lattice sways — first leaning one direction, then settling, then leaning the other way, in a steady breathing rhythm. The bars rock gently like a rope bridge swinging in the wind, each rung shifting just one step from center before swinging back. The top and bottom rails stay planted while the rows between them dance in a slow, repeating shimmy.

Grid: The dark void remains still. Only the lattice moves — its colored bars and the dark panes between them sway together as one body, like beads on a string sliding left or right. The frame never breaks apart or leaves the grid; it merely breathes sideways in place.

gpt-oss-120b — MARC valid, solved with 4 examples

Input

→

Prediction

vs

Expected

Reasoning (13607 chars)

We need to infer rule. Observations: Input shows a rectangular frame of colored cells (X, R, A, G) forming a thick border? Actually pattern: In example1, X forms a 4x? Let's view: Input has a block of X forming a 4x? Actually rows 2-11 have X pattern: looks like a 4x? Let's think: The pattern is a "window" with vertical bars? The description: lattice of panes divided by mullions, like a church window. The frame is rigid and geometric, every crossbar perfectly aligned. The wave passes bottom to top, each horizontal bar sways left/right one step then back, alternating direction each row? The output shows the X pattern shifted horizontally in a wave pattern: rows shift left or right relative to original.

Let's examine example1 more concretely. Input rows (0-index). Row2 (third row) has X X X X X X X X from col4 to col11 (8 X). Row3 has pattern: X . . X X . . X . . X (positions?). Let's index columns 0-14. Row2: cols4-11 X. Row3: col4 X, col5 ., col6 ., col7 X, col8 X, col9 ., col10 ., col11 X, col12 ., col13 ., col14 X? Wait there are 15 columns. Actually row3: ". . . . X . . X X . . X . . ." So X at col4, col7, col8, col11. Row4 same as row3. Row5: ". . . . X X X X X X X X . . ." same as row2. Row6: ". . . . X . . X X . . X . . ." same as row3. Row7 same. Row8 same as row2. Row9 same as row3. Row10 same as row3. Row11 same as row2.

So pattern is a grid of X forming a 4x? Actually it's like a 4x? The pattern resembles a "window" with vertical bars? But anyway.

Output: Row2 becomes ". . . X X X X X X X X . . .". So the block of X shifted left by 1 column (starting at col3). Row3 unchanged? Output row3: ". . . . X . . X X . . X . . ." same as input row3. Row4: ". . . . . X . . X X . . X . ." This is shifted right? Let's compare input row4 same as row3. Output row4 has X at col5? Actually output row4: ". . . . . X . . X X . . X . ." So X at col5, col8,9,12. That's like original row3 shifted right by 1? Original row3 had X at col4,7,8,11. Shift right gives col5,8,9,12 matches. So row4 shifted right.

Row5 (input row5 same as row2) output row5: ". . . . X X X X X X X X . . ." unchanged? Actually output row5 same as input row5 (starting at col4). So pattern: rows alternate shift left, none, right, none? Let's list:

Rows with full horizontal bar (row2,5,8,11) maybe shift left on row2, unchanged on row5? Wait row5 output same as input (col4). Row8 output ". . . . . X X X X X X X X . . ." shift right? Actually output row8: ". . . . . X X X X X X X X . . ." That's shift right by 1 (starting col5). Row11 output same as input? Row11 output ". . . . X X X X X X X X . . ." unchanged.

So pattern across rows: Starting from bottom? The description says wave from bottom to top, each horizontal bar sways first leaning one direction, then settling, then leaning other way, in a steady breathing rhythm. So maybe the rows of the lattice (horizontal bars) move left, then stay, then right, then stay, repeating as you go up.

Let's examine example2 with R pattern. Input

... [10703 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.69)

Prediction

Expected

Figurative + 4 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 100: "Finding the one bird that stayed with its flock while all ot ... [26 chars truncated]" k=1 alts: 0/15 1f85a75f

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input is a large grid (typically 30x30) with a black background. Several colors are scattered sparsely and seemingly randomly across the grid as noise pixels. Among all the colors present, exactly one color appears only in a small, compact, contiguous cluster rather than being scattered widely. This rare color forms a distinct shape, typically spanning only a few rows and columns, while every other non-black color appears many times distributed broadly across the entire grid.

Do: 1. For each non-black color in the grid, count how many disconnected pixels or small groups it forms and assess its spatial distribution. 2. Identify the one color whose pixels are all concentrated in a single small, compact region rather than scattered across the grid. This is the signal color; all other colors are noise. 3. Find the bounding box of all pixels of the signal color — the smallest rectangle that contains every pixel of that color. 4. Extract that bounding box from the grid, preserving the signal-color pixels in their positions and replacing all other cells (black or noise-colored) within the bounding box with black. 5. Output the extracted bounding box as the result.

Grid: The input grid is large (e.g., 30x30) with a black (.) background. The output grid is small, matching the dimensions of the bounding box of the signal-color cluster — typically between 3x3 and 5x5. The output contains only black (.) and the signal color.

Figurative Descriptions (16 variants)

"Finding the one bird that stayed with its flock while all others scattered in a storm."

120b: k=1

See: A vast dark night sky is filled with fireflies of various colors, each species flung far and wide by the wind — their lights blinking in lonely isolation across the expanse. But one species refused to scatter. Somewhere in the darkness, a small family of identically-colored lights huddles together, their glow forming a tiny constellation. They are the only ones who kept their formation while every other color drifted apart into solitary wandering.

Do: Search the sky for the loyal flock — the one color whose members never abandoned each other. Every other color is a diaspora, its individuals lost across the void. But this one color held hands. Once you find the huddled family, draw a frame around their gathering and lift it out of the darkness. Let the scattered loners fade away; only the constellation and the night around it remain.

Grid: The night sky is a large dark canvas sprinkled with colored sparks. The rescued constellation is a small portrait — a tight frame holding just the huddled family against the dark, no bigger than a few steps across.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (9735 chars)

We need to infer rule from example. Input grid large, output small 3x3? In example output is 3 rows, 3 columns: pattern of Y and . . Let's examine example: Input contains many colors: R, Y, etc. The description: find color whose members are contiguous (huddled) while others are scattered. Then output a tight frame around that cluster, with background black? Output shows only that cluster with its colors? In example output shows Y cluster shape: a plus? Let's map: Output:
. Y .
Y Y Y
. Y Y

That's shape of Y's cluster? In input, we need locate Y's that are together. Let's scan example input (given). Hard but we can approximate: There are Y's in rows 9-12 maybe. Let's locate Y positions: Row indices starting 0? Let's list rows with Y:

Row9 (0-index?) Actually line 10: ". . . . . . . . . . . Y . . . . . . . . . . R . . . . . . . ." (row9). Contains a single Y at column 11.

Row10: ". . . . . . . . . . . Y Y Y . . . . . . . . R . . . . . . . ." (row10) Y at columns 11,12,13.

Row11: ". . . . . . . . . . . . Y Y . . . . . . . . . . . . . . . . ." (row11) Y at columns 11,12.

Thus Y's form a cluster of shape maybe like a small block. Let's map coordinates: row9 col11; row10 col11-13; row11 col11-12. That's shape:

Row9: . Y .
Row10: Y Y Y
Row11: . Y Y

Exactly matches output (with . for black). So rule: find color that forms a connected component (4-connected?) that is larger than 1? But other colors also have multiple cells but scattered. Need to find the only color whose cells are all adjacent (forming a single connected component) while other colors have multiple components. In example, Y appears only in that cluster (connected). R appears many scattered solitary cells. So we extract bounding box of that component and output that subgrid, with other cells turned to . (black). So output is minimal bounding rectangle containing that component, preserving its colors (Y) and turning other cells within rectangle to . (if empty). Indeed bounding box rows 9-11, cols 11-13 gives 3x3 grid as output.

Thus rule: Identify the color that forms a single connected component (i.e., all its cells are connected) while all other colors have multiple components (i.e., are scattered). Then output the minimal bounding rectangle of that component, with that component's cells colored, others as . (black). If component includes other colors? Not.

Now apply to test input. Need to find which color appears as a single connected cluster (all its cells connected) while other colors appear in multiple disconnected groups. Let's parse test input grid. It's 30 rows? Let's count lines: Provided 30 lines? Let's count quickly.

I'll index rows 0-? Let's list each row with characters.

Row0: ". . B . G . A . . . . . . . . . . . . . G . A . . . . . . . ."
Row1: ". . A G . . . . . . . . . . . . . A . . A . . . . . . B B ."
Row2: "B . . . . . . . . . . . . . . . . . . A A A B G . . . . . . ."
Row3: ". . A . . . . . . A . . B . . . A . . . G . . G . . . . . . A"
Row4: "B . . . . . . . .

... [6735 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.01)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 136: "Musical chairs at a costume party — everyone moves one ... [43 chars truncated]" k=1 alts: 8/15 2601afb7

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The grid has an orange (O) background. Vertical bars of a single non-orange color occupy every other column (odd-indexed columns: 1, 3, 5, 7, 9, …). Each bar is anchored to the bottom row and extends upward for some height. Different bars have different colors and different heights. Even-indexed columns (0, 2, 4, …) are entirely orange and serve as separators.

Do: 1. Identify all vertical bars on odd-indexed columns. For each bar, record its color and its height (number of consecutive non-orange cells from the bottom up). 2. List the bars in order from left to right. 3. Rotate the sequence of colors one position to the RIGHT (the rightmost bar's color wraps around to become the leftmost bar's color). 4. Simultaneously rotate the sequence of heights one position to the LEFT (the leftmost bar's height wraps around to become the rightmost bar's height). 5. Redraw the grid: assign each bar position its new color and new height, filling from the bottom up, with all remaining cells orange.

Grid: The grid dimensions match the input exactly (same number of rows and columns). The background color is orange (O). Bars exist only on odd-indexed columns (1, 3, 5, …), and even-indexed columns remain entirely orange. Each bar is a contiguous vertical stripe of one color anchored to the bottom edge.

Figurative Descriptions (16 variants)

"Musical chairs at a costume party — everyone moves one seat right but leaves their shadow behind."

120b: k=1

See: Against a sunset-orange ballroom floor, costumed dancers stand in evenly spaced lanes, each wearing a single vivid color. Some dancers are tall, some short — their heights tell you how long they've been dancing. They stand rooted to the floor's edge, stretching upward like flames of different lengths, separated by empty aisles of orange light.

Do: The music stops and the dancers play a strange double game. Each dancer strips off their costume and tosses it one partner to the right — the dancer at the far end throws theirs all the way to the front of the line, like a boomerang. But their shadows — the silhouettes they cast on the wall — slide the opposite way, each shadow stretching to fill the space one partner to the left, with the frontmost shadow wrapping to the back. When the music resumes, each lane now wears a borrowed costume at a borrowed height: someone else's color, someone else's stature.

Grid: The ballroom keeps its exact dimensions. The orange floor fills every even lane and all the empty air above the dancers. The dancers only ever stand in the odd lanes, rooted to the bottom edge, rising upward in solid columns of color. After the swap, every lane still has exactly one costumed figure — but the colors and heights have been reshuffled as if the party spun like a carousel.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (13165 chars)

We need to infer rule from example. Input grid: 9x9? Actually example 1 input 9x9. Output also 9x9. Let's map.

Represent coordinates: rows top to bottom, columns left to right.

Input example1 grid (I'll index). Let's write rows:

1: O O O O O O O O O
2: O O O O O M O O O
3: O O O O O M O O O
4: O O O O O M O O O
5: O O O O O M O O O
6: O O O O O M O O O
7: O O A O M O B O
8: O W O A O M O B O
9: O W O A O M O B O

Wait rows 7-9 have fewer columns? Actually they have 9 columns: row7: O O A O M O B O (that's 8?) Let's count: "O O A O M O B O" = 8 entries. Something off. Maybe spaces misaligned. Let's reconstruct from description: The pattern likely: columns: 1-9. The input likely has orange floor in even columns (2,4,6,8) and black '.' in odd columns? But example uses O for orange floor, and other colors for dancers in odd columns. In example, there are columns with O and other colors.

Let's examine column pattern: In rows 2-6, there is a vertical column of M at column6? Actually row2: O O O O O M O O O => M at column6 (even). But dancers should be in odd columns only. However M appears in even column (6). Hmm maybe orange floor is even rows? Wait description: orange floor fills every even lane and all empty air above dancers. Dancers only stand in odd lanes, rooted to bottom edge, rising upward in solid columns of color. So lanes = columns? Even lanes (columns) are orange O throughout. Odd lanes have vertical columns of colors (maybe multiple colors stacked). In example, column6 is even? 6 is even, but it's M (magenta) not O. So maybe lanes are rows? Let's think: "lanes" could be rows. Even rows are orange floor. Dancers stand in odd rows, columns? But they are columns of color rising upward from bottom. So each dancer is a column of color in odd columns, with bottom at bottom row. In example, bottom rows have colors A, W, B etc in certain columns.

Let's examine bottom rows (rows 7-9). Row9: O W O A O M O B O. So there are colors W at column2? Actually column2 is W (maroon) which is odd column? column2 even. Hmm.

Maybe the grid is transposed: The dancers are rows, not columns. Let's think differently.

Better to analyze transformation: Output grid shows colors shifted right for costumes and left for shadows. Essentially each "dancer" column (odd lane) has a color column of some height (height = number of cells of that color from bottom up). The rule: shift the color columns right by one lane (wrapping) while shifting the empty orange floor left? Wait description: each dancer strips costume and tosses it one partner to the right — dancer at far end throws theirs to front of line (like boomerang). So costumes move right (to next odd lane?) The shadows slide opposite way left, each shadow stretching to fill space one partner to the left, with frontmost shadow wrapping to back. So the heights (shadows) shift left.

Thus after transformation, each odd lane still has a column of color, but the color (costume) is from previous lane (right n

... [10177 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.83)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 141: "A popularity contest where colors line up for a podium photo ... [50 chars truncated]" k=1 alts: 2/15 2753e76c

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black background with several colored rectangles scattered across it. Each rectangle is a solid block of a single color (e.g., green, yellow, azure, red, blue) and can be any size (2×2, 3×3, 4×4, 2×3, etc.). Multiple rectangles may share the same color but appear as separate, non-touching blocks at different positions in the grid.

Do: 1. Identify every distinct contiguous rectangle of non-black color in the input grid. Two blocks of the same color that do not touch are separate rectangles. 2. Count how many rectangles exist for each color. 3. Rank the colors from most rectangles to fewest. 4. Determine the output width as the maximum rectangle count across all colors. The output height equals the number of distinct colors present. 5. Build the output grid as a right-aligned staircase: the top row corresponds to the color with the most rectangles and is fully filled (width cells of that color). Each subsequent row corresponds to the next color in descending count order, with that many cells of its color right-justified in the row (padded with black on the left). The result is a staircase descending from upper-right to lower-left.

Grid: The output grid has width equal to the largest rectangle count among all colors and height equal to the number of distinct colors. The background is black (.). Each row contains one color, right-aligned, with the number of colored cells equal to that color's rectangle count. Rows are ordered top-to-bottom from highest count to lowest count.

Figurative Descriptions (16 variants)

"A popularity contest where colors line up for a podium photo, the most popular standing proudest at the front."

120b: k=1

See: Scattered across a dark field are colonies of colored creatures — some colors have established many small settlements, others only a few. Each settlement is a solid rectangular encampment, and even siblings of the same color keep to their own separate camps, never touching.

Do: A talent show is being judged. Each color is a contestant, and the number of separate encampments they've founded is their vote count. The host calls them to the stage in order of popularity — the color with the most fans steps up first, standing tall across the full width of the stage. Each subsequent contestant steps up behind, but shyly — they only occupy as many spotlights as votes they earned, hugging the right edge of the stage while darkness fills the empty seats to their left. The stage shrinks into a staircase descending from the star to the underdog.

Grid: The stage is only as wide as the highest vote count and only as tall as there are contestants. Each row is one color's moment in the spotlight — right-justified, with their earned spotlights glowing and the rest dark. The champion fills the top row completely; each act below is one step further from the edge, forming a dark triangle of empty seats in the upper-left corner.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (7119 chars)

We need to understand puzzle. Input grid contains colored rectangles (solid rectangular encampments) each color may have multiple separate rectangles, never touching. Need to count number of separate encampments per color (vote count). Then output grid is a staircase: rows correspond to colors sorted descending by count (most votes first). Width equals highest vote count. Each row is right-justified: fill with that color's count of cells on right side, left side '.' (black). So output grid dimensions: number of distinct colors = rows; columns = max count.

In example 1, we need verify.

Let's parse example 1 input (16x16). Need to count separate rectangular components per color.

Colors present: Y, R, G, A. Possibly others? Let's see.

We need to count number of separate rectangular encampments for each color.

From visual: Y appears in top rows: first row has Y Y at columns 9-10? Actually row1: ". . . . . . . . Y Y . . . R R .". So Y rectangle maybe 2x2? Let's map.

Better to count components: For each color, find distinct rectangular blocks not touching.

From example output: rows are R, G, A, Y (order). Output grid:

R R R R
. G G G
. . A A
. . . Y

Thus counts: R count = 4, G count =3, A count=2, Y count=1. So they had 4,3,2,1 separate rectangles respectively.

Let's verify with input: R appears many times; likely 4 separate rectangles. G appears 3 rectangles. A appears 2 rectangles. Y appears 1 rectangle.

Thus rule matches.

Now test input: need to count components per color.

Colors present: A, G, Y. Possibly others? Let's scan.

Rows 0-15 (16x16). I'll index rows.

Row0: ". . . . . . A A . . . . . . . ."
=> A rectangle at columns 6-7 (0-index). Possibly part of larger rectangle.

Row1: ". G G G . . A A . . . . . . . ."
=> G rectangle columns1-3, A rectangle columns6-7.

Row2: ". G G G . . . . . . G G . . A A"
=> G rectangle columns1-3 (continuation of previous? same rows 1-2, columns1-3 forms rectangle 2x3? Actually rows1-2 have G at cols1-3, so that's one rectangle. Also G at cols9-10? Wait row2: after some spaces, "G G" at columns? Let's count: positions: 0 '.';1 G;2 G;3 G;4 '.';5 '.';6 '.';7 '.';8 '.';9 '.';10 G;11 G;12 '.';13 '.';14 A;15 A. So another G rectangle at cols10-11 (size maybe 2x? rows2-3?). Row3 similar.

Row3: ". G G G . . . . . . G G . . A A"
=> same as row2.

Row4: ". . . . . . . . . . . . . . . ."
Row5: ". . . . Y Y Y . . . . . . . . ."
Row6: ". . . . Y Y Y . . Y Y . . . . ."
Row7: ". . . . Y Y Y . . Y Y . . . G G"
Row8: ". . . . . . . . . . . . . . G G"
Row9: ". . . . . Y Y Y . . . . . . . ."
Row10: ". . . . . Y Y Y . . . . . . . ."
Row11: ". G G . . Y Y Y . G G G G . . ."
Row12: ". G G . . . . . . G G G G . . ."
Row13: ". . . . Y Y . . . G G G G . . ."
Row14: ". . . . Y Y . . . G G G G . . ."
Row15: ". . . . . . . . . . . . . . . ."

Now need to identify separate rectangular components for each color.

Let's map coordinates (row, col). We'll list components.

Color A: appears rows0-1 columns6-7 (2x2 block

... [4173 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.20)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 151: "Lightning rods on a rooftop attract bolts from storms on the ... [9 chars truncated]" k=1 alts: 9/15 29700607

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black background with two key features: (1) A horizontal 'header segment' of exactly 3 consecutive colored pixels in row 0, and (2) scattered 'anchor' pixels elsewhere in the grid, each matching one of the header colors. Each anchor pixel appears at most once per header color. Some header colors may have no corresponding anchor pixel.

Do: For each of the 3 header colors, from left to right occupying columns L, M, R in row 0: (1) Find whether an anchor pixel of that color exists elsewhere in the grid. (2) If an anchor exists at position (row_a, col_a): fill that color's header column from row 0 down through row_a, then fill row_a horizontally from col_a to the header column (inclusive), connecting the anchor to the column. (3) If no anchor exists for that color: extend that color's header column from row 0 all the way down to the last row of the grid. All other cells remain black.

Grid: The grid dimensions match the input exactly (no resizing). The background is black (.). Row 0 contains a contiguous 3-pixel colored segment; the rest of the grid is black except for isolated anchor pixels. The coordinate system is row-major with (0,0) at the top-left.

Figurative Descriptions (16 variants)

"Lightning rods on a rooftop attract bolts from storms on the horizon."

120b: k=1

See: A rooftop runs along the top of the grid — three lightning rods stand side by side, each a different color. Scattered across the dark sky below, lone storm clouds drift, each glowing with the same color as one of the rods that calls to it.

Do: Each lightning rod reaches downward through the darkness, searching for its matching storm cloud. When a rod finds its cloud, a bolt of lightning strikes — the rod's color streams straight down from the rooftop and then arcs horizontally to meet the cloud, forging an L-shaped path of light. If a rod has no cloud to attract, its energy simply pours all the way down to the ground, an unbroken pillar of color with nowhere else to go.

Grid: The rooftop sits along the top edge — three adjacent colored rods against a pitch-black sky. The storm clouds are single colored sparks floating in the darkness. Lightning bolts are vertical then horizontal, always connecting rod to cloud by the shortest L-shaped route through the rod's own column.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (7239 chars)

We need to infer rule from example. Input grid 12x12? Let's count. Example input lines: each line has 12 entries. Yes 12x12.

In example, top row has "... X G A ..." three colored rods? Actually rooftop: three adjacent colored rods along top edge. In example input top row: ". . . . X G A . . . . ." So at columns 5,6,7 (0-index?) there are X (grey), G (green), A (azure). That's three adjacent colored cells. They are the rods.

Storm clouds are single colored sparks elsewhere: In example input there are X at row5 col1? Let's see: row5 (0-index row4) line: "X . . . . . . . . . . ." So X at column1. Also A at row6 col12? Actually row6 line: ". . . . . . . . . . . A" so A at column12. Also G at row9 line: "G . . . . . . . . . . ." So G at column1. So clouds of same colors as rods.

Now output shows L-shaped connections: For each rod, vertical line down from its column until reaching row of its cloud, then horizontal line to cloud column.

Let's verify: Rod X at column5 (index4). Cloud X at row5 col1 (index0). So vertical line from row1 to row5 in column5? Output shows column5 (X) from rows1-4? Let's examine output rows: Row1 same as input. Row2: ". . . . X G A . . . . ." So X appears at row2 col5. Row3 same. Row4 same. Row5: "X X X X X G A . . . . ." So row5 has X across columns1-5 (horizontal). So vertical line from rows1-5 column5 (X). Then horizontal from column5 to column1 (cloud) at row5. Yes.

Rod G at column6 (col6). Cloud G at row9 col1. Output: vertical G from rows1-8 column6? Let's see rows2-4 have G at col6? Row2: ". . . . X G A . . . . ." includes G at col6. Row3 same. Row4 same. Row5: "X X X X X G A . . . . ." includes G at col6. Row6: ". . . . . G A A A A A A" includes G at col6. Row7: ". . . . . G . . . . . ." includes G at col6. Row8: ". . . . . G . . . . . ." includes G at col6. Row9: "G G G G G G . . . . . ." includes G across columns1-6. So vertical line down to row9 col6, then horizontal left to col1.

Rod A at column7 (col7). Cloud A at row6 col12. Output: vertical A from rows1-5 col7? Row1 has A at col7. Row2 same. Row3 same. Row4 same. Row5 same. Row6: ". . . . . G A A A A A A" includes A at col7 and also A's across columns8-12 (horizontal). So vertical down to row6 col7, then horizontal right to col12.

If a rod has no cloud, vertical line goes all the way to bottom (ground). Not present in example.

Thus rule: Identify three adjacent colored cells on top row (rods). For each, find matching colored cell elsewhere (cloud). If found, draw vertical line of that color from rod down to the row of cloud (including rod row and cloud row). Then draw horizontal line from that column to cloud column on that row. If multiple clouds of same color? Probably only one.

If no cloud, draw vertical line down to bottom row (row12). No horizontal.

Now test input: Let's parse.

Grid 12x12.

Rows:

1: ". . . Y A X . . . . ."
2: ". . . . . . . . . . ."
3: ". . . . . . . . . . ."
4: ". . . . . . . . . . ."
5: ". . . . . . . . . . ."
6: ". . . . . . 

... [4248 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.92)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 194: "A curtain divides two dancers, and the spotlight finds whoev ... [16 chars truncated]" k=3 alts: 14/15 34b99a2b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains a single vertical column of yellow (Y) cells that divides the grid into two equal halves: a left half and a right half. The left half contains azure (A) cells and black (.) background cells. The right half contains grey (X) cells and black (.) background cells.

Do: Remove the yellow divider column and compare the left and right halves cell by cell. For each position, apply an XOR rule: if exactly one of the two sides has a colored cell (A on the left or X on the right) and the other side is black (.), output red (R). If both sides have a colored cell or both sides are black, output black (.). The result is a grid with only red (R) and black (.) cells.

Grid: The output grid has the same number of rows as the input and a width equal to half the input width minus the divider column. For example, a 5×9 input with a 1-cell-wide divider produces a 5×4 output. The background is black (.) and the only foreground color is red (R).

Figurative Descriptions (16 variants)

"A curtain divides two dancers, and the spotlight finds whoever dances alone."

120b: k=3

See: A golden curtain hangs down the center of a stage, splitting it into two wings. In the left wing, azure performers take their positions among empty squares. In the right wing, grey performers mirror the same stage layout. The black squares are darkness — empty stage floor.

Do: Raise the curtain and let the two troupes face each other. Wherever a performer on one side finds only darkness across from them — no partner, no mirror — a red spotlight ignites on that square, marking the lonesome dancer. But where two performers lock eyes across the divide, or where darkness meets darkness, the stage stays dark. The spotlight only loves the unmatched.

Grid: The curtain vanishes and the two wings collapse into one stage. What remains is a smaller floor plan lit only by red spotlights where solitary dancers stood, surrounded by the darkness of matched pairs and empty silence.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (6226 chars)

We need to infer rule. From description: golden curtain down center splitting into two wings. Left wing azure performers (A) among empty squares (.) . Right wing grey performers (X) mirror same stage layout. So input grid has left side with A and . and right side with X and .? Also Y appears maybe as something else? In examples, Y appears as a column maybe representing curtain? Let's examine Example1 input dimensions: 5 rows, 9 columns? Let's count first row: ". A . . Y . X X ." that's 9. So curtain column is Y? Yes Y column in middle. Left side columns before Y are left wing, right side after Y are right wing. The left wing contains A and . (azure and black). Right wing contains X and . (grey and black). Possibly also other colors but not in examples. The rule: compare each row left side pattern with right side pattern mirrored across curtain? "mirror the same stage layout" meaning right wing is mirror of left wing? Actually description: right wing performers mirror the same stage layout. So the pattern of A's on left side is mirrored as X's on right side? In input, we see X's placed where A's are mirrored? Let's test Example1: Row1 left of Y: ". A . ." positions columns 1-4. Right of Y: ". X X ." columns 6-9. Mirror of left across Y would be reversed order: columns 1-4 reversed = ". . A ." (since left pattern . A . . reversed is . . A .). Right side actual: . X X . . Wait there are 4 columns after Y? Actually after Y there are ". X X ." that's 4 columns. Mirror of left: . . A . ; but we have . X X . . Not matching. Let's map: maybe X corresponds to A positions but not mirrored? Let's list left A positions: column2. Right side X positions: columns7 and8? Actually columns after Y: col6 '.' col7 'X' col8 'X' col9 '.' So X at col7 and8. Not matching.

Maybe rule: For each row, compare left side A pattern with right side X pattern directly (same column offset from curtain). If left has A and right has X at same offset, they are matched pair and become dark (.) in output. If left has A but right has . (no partner) then output red at that left position? But output grid is smaller (collapsed). Output dimensions are 5 rows, 4 columns (since curtain removed and wings collapsed). Indeed output examples have 4 columns. So they collapse left and right wings into one by overlaying: For each row, compare left side cell and right side cell (mirrored?) and produce a single cell in output: If exactly one of them is a performer (A or X) and the other is empty (.), output R (red). If both are performers (A and X) or both empty, output . (black). That matches description: "where a performer on one side finds only darkness across from them — no partner, no mirror — a red spotlight ignites on that square". So output cell corresponds to that position (maybe left side position). The output grid size equals number of rows x number of columns left of curtain (or right). In examples, left side columns count = 4 (since total 9 columns, Y at column5). Output col

... [3364 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 207: "A talent show where the most popular act wins the encore." k=1 alts: 9/15 39a8645d

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains several small 3×3 shapes scattered on a black background. Each shape is a distinct pattern of colored cells within a 3×3 bounding box, with the remaining cells black. Multiple colors may be present (e.g., red, green, azure, blue, yellow, magenta), and some shape patterns appear more than once as exact duplicates — same color and same spatial arrangement of colored cells within their 3×3 bounding box.

Do: 1. Scan the input grid and identify every distinct 3×3 shape by extracting each connected or nearby cluster of non-black cells and reading its 3×3 bounding box pattern. 2. Group these shapes by their exact pattern (same color and same cell arrangement within the 3×3 box). 3. Count how many identical copies exist for each distinct shape pattern. 4. Select the shape pattern that has the highest number of identical copies in the grid. 5. Output that 3×3 pattern as the result.

Grid: The input is a rectangular grid (typically 14×14) with a black (.) background. The output is always a 3×3 grid showing the extracted shape pattern on a black background. Each shape occupies exactly a 3×3 bounding box within the input grid.

Figurative Descriptions (16 variants)

"A talent show where the most popular act wins the encore."

120b: k=1

See: A dark stage is scattered with small troupes of performers, each wearing matching costumes of a single color — blue dancers, red acrobats, yellow jugglers, azure mimes. Each troupe strikes a unique pose within its own small spotlight. Some acts are one-of-a-kind, but others have sent identical understudy groups who mirror the original's pose and costume perfectly, scattered elsewhere across the stage.

Do: The audience votes by recognition: whichever act has sent the most identical understudy copies of itself onto the stage is clearly the crowd favorite. That winning troupe — the one whose exact pose and costume appears more times than any other — earns the encore. Pull them into the spotlight alone and let the rest of the stage go dark.

Grid: The stage is a wide dark floor where each troupe occupies a small spotlight just large enough for a few performers. The encore spotlight shrinks to show only the winning troupe's formation against the darkness, exactly as they stood during the show.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (13543 chars)

We need to infer rule from example. Input grid large, output small 3x3? In example output is 3x3. Likely we need to find the most frequent pattern (shape) of a color cluster (connected component) that appears multiple times identical (same shape and color). Then output that shape isolated on minimal bounding box, with other cells empty (.)? In example, output shows Y . . etc. So winning troupe is Y (yellow). Let's examine example input to see clusters of Y.

We need to parse example input (14x15?). Let's map coordinates. But maybe easier: Identify distinct shapes of each color. Count identical copies (same shape orientation). The most copies wins. Then output that shape in its original orientation, placed in top-left? In output they placed shape as appears (maybe bounding box trimmed). For Y shape, output grid 3x3 with Y pattern.

Let's examine Y positions in example input. Write rows (0-index). I'll transcribe quickly.

Row0: ". . . . . . . B . B . . . ."
Row1: ". . Y . . . . B B B . . . ."
Row2: ". . . Y Y . . . B . . . . ."
Row3: ". . Y . . . . . . . . Y . ."
Row4: ". . . . . . . . . . . . Y Y"
Row5: ". . . . . . R . R . . Y . ."
Row6: ". B . B . . . R . . . . . ."
Row7: ". B B B . . R . R . . . . ."
Row8: ". . B . . . . . . Y . . . ."
Row9: ". . . . . . . . . . Y Y . ."
Row10: ". . Y . . . . . . Y . . . ."
Row11: ". . . Y Y . . . . . . B . B"
Row12: ". . Y . . . . . . . . B B B"
Row13: ". . . . . . . . . . . . B ."

We need to find Y clusters. Let's locate each Y coordinate.

Row1 col2 (0-index). Row2 col3,4. Row3 col2, col11. Row4 col12,13. Row5 col10. Row8 col9. Row9 col10,11. Row10 col2, col9. Row11 col3,4. Row12 col2. That's many Y cells, many separate clusters.

We need to find identical shapes. Likely there are multiple copies of a small shape of Y of size maybe 3 cells? Let's examine clusters.

Cluster A: Row1 col2 alone? Check adjacency orthogonal? Usually connectivity maybe 4-direction. Row1 col2 adjacent to any other Y? Row2 col3 is diagonal, not orthogonal. So isolated single Y.

Cluster B: Row2 col3 and col4 adjacent horizontally (size 2 line). Also maybe adjacent to row1 col2 diagonal not count. So shape of two horizontal Y.

Cluster C: Row3 col2 isolated? Row3 col11 isolated.

Cluster D: Row4 col12,13 adjacent horizontally (size2). Row5 col10 isolated. Row8 col9 isolated. Row9 col10,11 adjacent horizontally (size2). Row10 col2 isolated, col9 isolated. Row11 col3,4 adjacent horizontally (size2). Row12 col2 isolated.

Thus we have many size1 and size2 horizontal pairs. Which shape appears most? Horizontal pair appears many times: at (row2 col3-4), (row4 col12-13), (row9 col10-11), (row11 col3-4). That's 4 occurrences. Single Y appears many times: count singles: row1 col2, row3 col2, row3 col11, row5 col10, row8 col9, row10 col2, row10 col9, row12 col2 = 8 singles. So singles more frequent. But output shows a shape of Y with pattern:

Output:
Y . .
. Y Y
Y . .

That's a 3x3 shape with Y at (0,0), (1,1),(1,2),(2,0). That's like an L

... [10588 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.33)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 218: "A living cell divides, turning itself inside-out and stretch ... [59 chars truncated]" k=2 alts: 6/15 3befdf3e

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black background containing one or more rectangular concentric two-color patterns. Each pattern consists of an outer frame of one color (the 'border color') surrounding a filled rectangular interior of a second color (the 'core color'). The core region can be any rectangular dimensions (1×1, 1×2, 2×2, etc.). The border is always exactly one cell thick. Multiple patterns may appear on the same grid at different positions and do not overlap.

Do: For each pattern in the input, perform the following steps: 1. Identify the border color (B) and core color (C), and note the dimensions of the core region (height h, width w). 2. Color-swap the original pattern to create a new center: replace every B cell with C, and every C cell with B. This produces a C-frame surrounding a B-filled interior, positioned exactly where the original pattern was. 3. Extend the border outward by adding rectangular bands of B cells in all four cardinal directions. Each band has thickness equal to the core dimensions along the perpendicular axis: - Top and bottom bands: extend h rows outward, spanning the full width of the original pattern. - Left and right bands: extend w columns outward, spanning the full height of the original pattern. 4. The extensions do not include diagonal corners — only the cardinal-direction projections of the original pattern's footprint are filled. 5. All cells not covered by any pattern or its extensions remain black.

Grid: The grid dimensions match the input. The background is black (.). Patterns and their extensions must fit within the grid boundaries. Each pattern is processed independently; if multiple patterns exist, they do not interact or overlap.

Figurative Descriptions (16 variants)

"A living cell divides, turning itself inside-out and stretching its membrane outward like a starfish reaching its arms."

120b: k=2

See: On a dark ocean floor, you find small living cells — each one a nucleus of one color wrapped snugly in a membrane of another. They sit quietly, compact and self-contained, like seeds waiting to germinate.

Do: Each cell undergoes mitosis in place. First, the cell inverts — the membrane sinks inward to become the new nucleus, while the old nucleus material rises to coat the surface, as if the organism turned itself inside-out. Then, energized by this transformation, the creature reaches outward in the four cardinal directions, projecting thick pseudopods of its old membrane color. Each pseudopod's thickness mirrors the proportions of the original nucleus — a wide nucleus sends out wide arms, a tall nucleus sends out tall arms. The creature stretches only along the compass lines, never diagonally — like a starfish that grew exactly four arms aligned to the grid.

Grid: The ocean floor stays dark wherever no creature has claimed territory. Each organism's body and outstretched arms must remain within the boundaries of the world. Multiple creatures, if present, grow independently without touching one another.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (14155 chars)

We need to understand rule. From description: each "cell" is a nucleus of one color wrapped in membrane of another. In examples, input shows a shape: a 3x3 block? Let's examine Example1 input: there is a 4x4? Actually coordinates: rows 3-6 columns 3-6? Let's map.

Input example1: rows index 0-9. At row3 col3-6: R R R R (a horizontal line). Row4 col3 R, col4 O, col5 O, col6 R. Row5 same. Row6 col3-6 R R R R. So shape is a 4x4 square with outer border R and interior O (2x2). So outer color = membrane (R), inner color = nucleus (O). That's a "cell". Output shows transformation: The outer border becomes interior? Let's see output: rows 1-8? Output row1 (index1) has R R R R at col3-6 (same as input row3). Row2 same. Row3: at col2-7: R R O O O O R R (a larger shape). Row4: R R O R R O R R. Row5 same. Row6: R R O O O O R R. Row7: row7 col3-6 R R R R. Row8 same. So the original 4x4 cell expanded outward with arms of membrane color (R) extending outward in four directions, thickness equal to original nucleus dimensions (which is 2x2). Indeed original nucleus O is 2x2. So arms thickness = 2 cells wide/tall. The membrane color R becomes new nucleus (fills original interior?) Wait after transformation, the "membrane sinks inward to become new nucleus, while old nucleus rises to coat surface". So original membrane color (R) becomes interior (new nucleus) and original nucleus color (O) becomes outer border (membrane). In output, the central area (original 4x4) now has R interior? Let's see output rows3-6 columns3-6: they are R R R R? Actually row3 (index3) col3-6 are O O O O? Wait row3 is ". R R O O O O R R .". At col3 (0-index) maybe? Let's compute: row3 (index3) has pattern: . R R O O O O R R . .? Actually output row3: ". R R O O O O R R .". So columns: 0 '.'; 1 'R';2 'R';3 'O';4 'O';5 'O';6 'O';7 'R';8 'R';9 '.' . So the original 4x4 region (cols3-6 rows3-6) now contains O's (the original nucleus) as interior? Wait original membrane was R, now interior is O? Actually row3 col3-6 are O O O O (original nucleus). So yes original nucleus O becomes interior (new nucleus). The original membrane R becomes outer border (including arms). So rule: swap colors: membrane becomes outer, nucleus becomes inner.

Arms: extend outward from the new outer border (original membrane color) in four cardinal directions, thickness equal to dimensions of original nucleus (height and width). In example1, original nucleus is 2x2, so arms are 2 cells thick. Indeed arms extend 2 cells outward from each side of the original square, forming a larger plus shape.

Specifically, the output shape is a larger square of size (original size + 2*thickness)?? Let's compute: original outer size = 4 (including membrane). Thickness = 2. Output outer size = 8? Actually outermost R border extends 2 cells beyond original border on each side, making total size 8? Let's see: original outer border at rows3-6, cols3-6. Output outermost R border appears at rows1-8, cols1-8? Row1 (index1) has R at cols3-6 on

... [11155 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.19)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 223: "Scattered seeds fall and pile into pyramids on the ground." k=1 alts: 5/15 3d6c6e23

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black background with one or more columns containing colored cells. Each such column has colored cells scattered vertically in its upper portion, separated by gaps of black cells, with empty space below. A column may contain cells of a single color or multiple colors. When multiple colors appear in a column, they are grouped contiguously from top to bottom (e.g., all red cells appear above all blue cells). The total number of colored cells in each column is always a perfect square (1, 4, 9, 16, ...).

Do: For each column that contains colored cells, do the following independently: (1) Count the total number of colored cells N in that column. Compute the triangle size as sqrt(N), call it S. (2) Record the colors from top to bottom in the column, preserving their order. (3) Clear the entire column. (4) Build a triangle of S rows at the very bottom of the grid, centered on that column. The bottom row of the triangle sits on the last row of the grid. Row widths from top to bottom are 1, 3, 5, ..., 2S-1 (odd numbers). Each row is horizontally centered on the column. (5) Fill the triangle rows with colors from top (narrowest) to bottom (widest). Assign colors in the same top-to-bottom order as they appeared in the input column: the first color fills cells starting from the top row of the triangle, consuming one cell per triangle position left-to-right, top-to-bottom, until that color's count is exhausted, then continue with the next color. All cells not part of any triangle become black.

Grid: The output grid has the same dimensions as the input grid. The background is entirely black except for the triangle(s) drawn at the bottom. Multiple columns may each produce their own independent triangle, and these triangles coexist on the grid without overlapping. Triangles are always bottom-aligned (touching the last row) and horizontally centered on their source column.

Figurative Descriptions (16 variants)

"Scattered seeds fall and pile into pyramids on the ground."

120b: k=1

See: Against a night sky, colored seeds float in vertical columns — hovering in the upper air, spaced unevenly as if suspended by invisible threads. Some columns hold seeds of a single hue; others carry layers of different colors stacked like geological strata, one color giving way to the next below it. The ground below sits empty, waiting.

Do: Gravity finally switches on. Every floating seed in a column falls, but rather than simply dropping straight down, the seeds accumulate into a neat triangular mound resting on the ground floor — like sand pouring through an hourglass and forming a perfect pile. The colors preserve their original layering: whatever was highest in the sky settles at the peak of the mound, and the deepest colors spread across its wide base. The more seeds that fall, the taller and broader the pyramid they build.

Grid: The sky clears completely — all black — and only the triangular mounds remain, each sitting flush against the bottom edge of the grid, centered beneath where its seeds once floated. If multiple columns had seeds, multiple independent pyramids form side by side, each loyal to its own column. The night reclaims everything the pyramids don't touch.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (9262 chars)

We need to understand the puzzle. The description: seeds floating in vertical columns, then gravity makes them fall forming triangular mound (pyramid) with colors preserving layering: highest in sky becomes peak, deepest colors spread across base. The output shows black (.) everywhere except pyramids at bottom.

We need to infer rule from example.

Example 1 Input: grid size? Let's count rows: appears 16 rows? Actually lines: Input shows 16 rows? Let's count: lines from ". . . . . . . . . . . . ." repeated. I see 16 rows maybe. Output also 16 rows. In input, there is column with O's at column 7? Let's map.

Let's index columns 0-11? Actually each row has 12 entries. So grid 16x12.

Input: many O's in column 6 (0-index). Let's list rows with O:

Row1 (index0): all .
Row2 (1): column6 O (since . . . . . . O . . . . .) yes.
Row3 (2): column6 O
Row4 (3): all .
Row5 (4): column6 O
Row6 (5): column6 O
Row7 (6): column6 O
Row8 (7): all .
Row9 (8): column6 O
Row10 (9): all .
Row11 (10): column6 O
Row12 (11): column6 O
Row13 (12): column6 O
Row14 (13): all .
Row15 (14): all .
Row16 (15): all .

So there are many O's in same column, but not contiguous; they are spaced.

Output: bottom rows have pyramid of O's centered at column6? Let's see output rows bottom three rows:

Row13 (index12): all .
Row14 (13): . . . . . . O . . . . . (single O at column6)
Row15 (14): . . . . . O O O . . . . (three O's columns5-7)
Row16 (15): . . . . O O O O O . . . (five O's columns4-8)

So pyramid of height 3 (rows 13-15) with base width 5. The number of O's originally? Count O's in column6: let's count: rows with O: indices1,2,4,5,6,8,10,11,12 = 9 O's. Pyramid of height 3 uses 1+3+5=9 cells. Yes matches.

Thus rule: For each column that contains colored cells (non-black), we take all those cells (any colors) and drop them to bottom forming a centered isosceles triangle (pyramid) of that color? Wait colors preserve layering: highest in sky becomes peak, deepest colors spread across base. In example only one color O, so all O.

If multiple colors in same column, we need to stack them in order: topmost becomes apex, then next rows of triangle? But description says colors preserve layering: whatever was highest in the sky settles at the peak of the mound, and the deepest colors spread across its wide base. So within a column, there may be multiple colors stacked vertically (maybe different colors). When they fall, they form a triangle where each row of triangle corresponds to a layer of colors? Actually need to think: Suppose column has colors from top to bottom: C1 (top), C2, C3 (bottom). When they fall, they will form triangle of height h where total number of cells = n (number of seeds). The triangle rows have widths 1,3,5,... up to (2h-1). The colors need to fill the triangle such that the apex cell is the topmost original color, and the base cells are the bottommost original colors. But there are more cells than original seeds? No, number of cells equals number of seeds, so 

... [6268 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 227: "Russian nesting dolls carved from colored sticks" k=1 alts: 7/15 3ee1011a

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black background containing several colored line segments (horizontal or vertical). Each segment is a contiguous run of a single non-black color, either all in one row or all in one column. The segments vary in length — some are just a single cell, others span multiple cells. No two segments share the same color.

Do: 1. Identify all distinct colored line segments in the input grid. For each segment, record its color and its length (number of cells). 2. Sort the segments by length in descending order (longest first). 3. The longest segment has length L. Create an L×L output grid. 4. Fill the output grid in layers from outside to inside: the longest segment's color fills the entire grid as the outermost border ring. The second-longest segment's color fills the next inner rectangle (inset by 1 on all sides). Continue for each subsequent segment, each forming the next inner ring, inset by 1 additional cell on each side from the previous. 5. The shortest segment (which may be a single cell) fills the center of the grid. If a segment has length 1, it fills only the single center cell (for odd-sized grids) or a 2×2 center block is not needed — each layer's thickness is exactly 1 cell on each side.

Grid: The output grid is a square with side length equal to the length of the longest segment found in the input. It contains concentric rectangular rings of color, one ring per segment, with no black background remaining. The outermost ring is 1 cell thick on each side, as is each subsequent inner ring.

Figurative Descriptions (16 variants)

"Russian nesting dolls carved from colored sticks"

120b: k=1

See: Scattered across a vast dark field lie colored sticks of different lengths — each a single pure color, each lying perfectly straight, no two alike. They look abandoned, but each one remembers its size.

Do: The longest stick claims authority and bends itself into a square frame — the outermost shell of a nesting doll. The next longest stick, humbled, curls inside that frame to form a smaller shell within. Each remaining stick, from longest to shortest, takes its turn nestling inside the one before it, like matryoshka dolls stacking inward. The tiniest stick — perhaps just a splinter — finds itself cradled at the very heart.

Grid: The result is a perfect square portrait of nested shells, each exactly one layer thick, with the mightiest stick's color as the outer skin and the smallest stick's color glowing at the core. No darkness remains — every inch is claimed by a shell.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (5338 chars)

We need to infer rule. From example: Input grid has colored sticks (lines) of single colors, each straight line (vertical/horizontal?) In example1 input: we see B vertical line? Let's examine: grid 21x21? Input shows B column at column? Let's map coordinates: rows 0-20. B appears at column 13? Actually line: ". . . . . . . . . . . . . . . . . . . . . ." many rows. At row2: "... B ..." at column? Count: there are 13 dots then B then rest. So B at column 13 (0-index). Also G vertical line at column 4? rows 3-6 have G at column4. M vertical line at column8? rows8-9 have M at column8. So three sticks: B longest? length? B appears from row2 to row7? Actually B appears rows2-7 (6 rows). G appears rows3-6 (4 rows). M appears rows8-9 (2 rows). So lengths: B=6, G=4, M=2. Output is a 6x6 square with nested layers: outermost B, then G, then M at center 2x2. So rule: take sticks lengths, sort descending, create square of size equal to longest length? Actually longest length =6 => output 6x6. Outer border color = longest stick's color (B). Next inner border (one cell thick) = second longest (G). Next inner border = third (M). Since lengths are even? The innermost region size = length of shortest? For lengths 2 => inner 2x2 region of M. Works.

Thus transformation: Identify each colored line (stick) in input: they are straight lines (vertical or horizontal?) In example they are vertical. Could be any orientation but likely vertical/horizontal. Determine length (number of cells). Sort descending. Create a square of size equal to longest length. Fill layers from outermost to innermost with colors according to descending lengths. If there are more sticks than can fit? Probably each layer thickness is 1 cell. The innermost region will be size = length of shortest stick (maybe also 1 cell if odd?). For odd lengths, inner region may be 1x1.

Now test input: Need to find sticks.

Let's parse grid: size? appears 22 columns? Let's count: first row has 22 entries? ". . . . . . . . . . . . . . . . . . . . . . ." that's 22. rows 0-20? Actually there are 22 rows? Let's count lines: Provided 22 rows (0-21). Let's index.

I'll list rows with coordinates.

Row0: all .
Row1: all .
Row2: all .
Row3: all .
Row4: ". . . . . . G . . . . . . . . . . . . . . . ." => G at column6? Count: columns 0-5 are ., column6 is G, rest . So G vertical line length? Let's see further rows.

Row5: all .
Row6: ". . . . . . . . . . O . . . . . . . . . . . ." => O at column10? Actually count: columns0-9 ., column10 O, rest .
Row7: same O at column10.
Row8: same O at column10.
Row9: all .
Row10: all .
Row11: ". . M M M M M . . . . . . . . . . . . . . ." => M from column2 to column6 (5 cells) horizontally? Actually row11 has M M M M M (5) starting at col2? Let's count: col0 ., col1 ., col2 M, col3 M, col4 M, col5 M, col6 M, col7 ., rest .
Row12: ". . . . . . . . . . A . . . . . . . . . . . ." => A at column10? Wait column10? Count: col0-9 ., col10 A, rest .
Rows13-18: similar A vertic

... [2434 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.25)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 229: "A treasure chest lies buried in a field of scattered gems &# ... [40 chars truncated]" k=2 alts: 10/15 3f7978a0

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains a rectangular frame made of three distinct elements: (1) two vertical columns of grey (X) cells forming the left and right walls, (2) azure (A) cells at the four corners connecting the tops and bottoms of the X walls, and (3) a row of A cells along the top edge and bottom edge between the corner A's. Inside the frame, some cells are azure (A) and the rest are black (.). Scattered across the grid outside this frame are additional azure (A) cells that serve as distractors.

Do: 1. Locate the two vertical columns of grey (X) cells in the grid. These form the left and right walls of a rectangle. 2. Identify the azure (A) cells that sit directly above and below each X column — these are the corner cells of the rectangle. The top row of the rectangle is the row containing the A corners above the X walls; the bottom row is the row containing the A corners below the X walls. 3. Extract the sub-grid defined by this rectangle: from the top corner row to the bottom corner row (inclusive), and from the left X column to the right X column (inclusive). 4. Output this extracted sub-grid exactly as it appears. All content outside the rectangle is discarded.

Grid: The input grid is variable in size with a black (.) background. Embedded in it is a rectangular frame whose width equals the horizontal distance between the two X columns (inclusive) and whose height equals the vertical span from the top A corners to the bottom A corners (inclusive). The output grid has exactly these dimensions. The frame's walls are grey (X), its corners and top/bottom edges are azure (A), its interior contains a mix of black (.) and azure (A) cells, and the background outside the frame is black with scattered azure distractors.

Figurative Descriptions (16 variants)

"A treasure chest lies buried in a field of scattered gems — only the iron-bound chest matters."

120b: k=2

See: Across a dark field, azure gemstones lie scattered like fallen stars. But among them, two pillars of grey iron stand upright, and azure cornerstone jewels cap their tops and bottoms, forming the frame of a buried chest. Inside the chest, some compartments hold azure gems while others sit empty. The gems strewn outside the chest are fool's gold — glittering distractions left by wanderers.

Do: Find the two iron pillars rising from the darkness — they are the sides of the chest. Look for the azure cornerstones that seal each pillar's top and bottom; these mark where the lid begins and the base ends. Lift the chest out of the ground whole, leaving every scattered gem behind in the dirt. What you hold in your hands is the treasure — nothing more, nothing less.

Grid: The dark field stretches in all directions, littered with stray azure gems that catch the eye. The chest is defined by its grey iron walls running vertically and its azure cornerstones and trim running horizontally across the top and bottom. Inside the chest, azure gems sit in some slots while others are empty black velvet. The chest's dimensions are fixed by its own construction — pillar to pillar, cornerstone to cornerstone — and everything beyond its edges is just empty earth.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (10614 chars)

We need to infer rule. From examples, we have input larger grid with many colors, output smaller grid focusing on a "chest" defined by grey pillars (X) and azure corners (A). The output seems to be a subgrid extracted: the rectangle bounded by the outermost X columns (vertical pillars) and outermost A rows (top and bottom). Inside that rectangle, we keep the original colors? Let's examine Example1.

Input1 grid size? Let's write rows (10? Actually given 9 rows? Let's count). Input lines:

1 ". A . . . . . . . . ."
2 ". . A . . . A . . . A"
3 ". . . . . . . . . A A"
4 ". . A . . . . . A . ."
5 "A . X . . . . . X . ."
6 ". . X . . A A . X . ."
7 ". . X . . . . . X . ."
8 ". . A . A . . . A . ."
9 ". A . . . . . . A A ."

So 9 rows, 11 columns.

Output:
A . . . . . A
X . . . . . X
X . . A A . X
X . . . . . X
A . A . . . A

That's 5 rows, 7 columns.

Now locate X pillars: In input, there are X's forming vertical lines? Let's map coordinates (row, col). Row5 col3 is X. Row6 col3 X, row7 col3 X, row8 col3? row8 col3 A not X. Row9 col3 . So vertical line at col3 rows5-7. Also there is another vertical line at col9? Row5 col9 X, row6 col9 X, row7 col9 X, row8 col9 A? Actually row8 col9 A, row9 col9 A? Wait row9 col9 A? Row9: ". A . . . . . . A A ." columns: 1 .,2 A,3 .,4 .,5 .,6 .,7 .,8 .,9 A,10 A,11 . So col9 A. So X column at col9 rows5-7. So two pillars of X at columns 3 and 9 rows5-7. The top and bottom corners are A's at the ends of these pillars: At top of pillar (row5) there is A at col1? Actually row5 col1 A, not at pillar column. But description: azure cornerstones cap their tops and bottoms. In example, top of each pillar maybe row5? Actually pillars start at row5 (first X). The top corner A maybe at same row but left/right of pillar? Let's examine output rectangle: It includes rows5-9? Output rows correspond to rows5-9? Output first row is "A . . . . . A". That matches row5 columns 1? Let's extract subgrid rows5-9 columns1-7? Row5 columns1-7: "A . X . . . ." => A . X . . . . Not match output first row (A . . . . . A). So not that.

Maybe rectangle defined by outermost X columns (col3 and col9) and outermost A rows (topmost A in those columns? Actually topmost A in column3? There is no A in column3. But there are A's at row5 col1 and row9 col2 etc.

Let's examine output content: It contains A's at corners: top-left A, top-right A, bottom-left A, bottom-right A. Also X's forming vertical sides inside? Output rows show X at column1 rows2-4? Actually output grid:

Row1: A . . . . . A
Row2: X . . . . . X
Row3: X . . A A . X
Row4: X . . . . . X
Row5: A . A . . . A

So pattern: rectangle of width 7, height5. The leftmost column: A, X, X, X, A. Rightmost column: A, X, X, X, A. So vertical sides are X except corners A. Inside there are some A's at row3 col4-5 (two A's) and row5 col3 A.

Now locate these in original input: The rectangle likely corresponds to region bounded by the two X pillars (col3 and col9) and the topmost A above left pillar and b

... [7680 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.28)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 236: "A flag unfurling in the wind, anchored at its bottom edge" k=1 alts: 15/15 423a55dc

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: The input grid has a black background with a single irregular closed shape made of one non-black color. The shape resembles a rough rectangle or polygon outlined in colored pixels, with a hollow interior of black pixels. The shape has a clearly identifiable bottom row (its lowest horizontal edge) and a top row (its highest horizontal edge).

Do: Identify the bottom row of the colored shape — this is the anchor row and does not move. For every other row of the shape, count how many rows above the bottom row it is; call this distance d. Shift all colored pixels in that row exactly d cells to the left. Pixels that would shift to negative column indices are discarded. All other cells remain black. The vertical positions of all rows stay the same; only horizontal positions change.

Grid: The grid dimensions remain identical to the input. The background is black (.). Only one non-black color is present. The shape occupies a contiguous vertical span of rows, and the transformation only affects columns within those rows.

Figurative Descriptions (16 variants)

"A flag unfurling in the wind, anchored at its bottom edge"

120b: k=1

See: Against a night sky, a colored banner hangs in stillness — its shape an irregular pennant stitched from a single hue, hollow at its heart like a window frame. The banner's bottom hem is nailed firmly to an invisible pole.

Do: A steady wind blows from the right. The bottom hem, nailed down, holds fast. But every row of fabric above it catches the breeze — and the higher it flutters from the anchor, the further the wind carries it leftward. Each row sails left by exactly as far as it stands above the grounded hem, as though the wind grows stronger with altitude.

Grid: The night sky keeps its darkness and its dimensions. Only the single-colored banner moves — its hem stays planted while the rest streams leftward like a flag caught in a rising gale.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6950 chars)

We need to understand transformation. The description: a banner shape irregular pennant stitched from a single hue, hollow at its heart like a window frame. Bottom hem is nailed to invisible pole. Wind from right, bottom hem stays, each row above moves left by exactly as far as it stands above the grounded hem (i.e., distance from bottom row). So basically shift each row left by number of rows from bottom of shape? Let's examine example.

Input shape: color R forms a shape like a hollow rectangle? Let's map coordinates. Input grid 11 rows, 13 columns? Actually given 13 columns. Let's index rows 0-10.

Rows with R:

Row3: columns 4-8 (5 Rs)
Row4: col4 R, col8 R (others .)
Row5: col4-5 R,R, col8 R
Row6: col5 R, col8-9 R,R
Row7: col5 R, col8 R
Row8: col5-9 R (5 Rs)

So shape is like a flag with bottom row row8 (index 8) having 5 Rs. The bottom hem is row8? Actually bottom of shape is row8 (since rows 0-2 empty, row9-10 empty). The bottom hem is anchored, not moving. Rows above shift left by distance equal to vertical distance from bottom row? Let's test.

Output rows:

Row3 (original row3) becomes shifted left by? In output row3 (index3) we have R R R R . . . . . . . . . (4 Rs at cols0-3). Original row3 had Rs at cols4-8 (5 Rs). After shift left by 4? Actually bottom row is row8. Distance from row3 to row8 = 5 rows (8-3=5). If shift left by 5, original cols4-8 shift to -1 to 3 (some off-grid). But output shows Rs at 0-3 (4 Rs). Not matching.

Maybe shift left by number of rows above bottom hem (i.e., row index relative to bottom of shape's top? Let's compute shift for each row based on its row number relative to bottom row of shape (row8). Row8 stays same (no shift). Row7 shift left by 1? Output row7 (index7) has . . . . R . . . R . . . . . Let's see original row7 had Rs at col5 and col8. After shift left by 1, they'd be at col4 and col7. Output row7 shows R at col4? Actually output row7: ". . . . R . . . R . . . . ." That's R at col4 and col8? Wait count: positions 0-12. Let's list: 0.,1.,2.,3.,4R,5.,6.,7.,8R,9.,10.,11.,12. So Rs at col4 and col8. Original had col5 and col8. So left shift of 1 moves col5->4, col8->7 but output shows col8 unchanged. So not simple shift.

Maybe the shape is being "sheared" such that each row is shifted left by its distance from bottom *but also the shape is hollow, so interior spaces become filled? Let's examine row6 (index6). Original row6: col5 R, col8-9 R,R. Output row6: ". . . R . . R R . . . . ." That's Rs at col3, col6, col7. Original col5->3 shift 2? col8->6 shift2, col9->7 shift2. So shift left by 2 for row6 (distance from bottom row8 is 2). That matches: row6 distance = 8-6=2, shift left 2. Good.

Row5 (index5) distance 3. Original row5: col4-5 R,R, col8 R. Shift left 3 => col1-2, col5. Output row5: ". R R . . R . . . . . . ." That's Rs at col1,2,5. Yes matches! Row4 distance 4. Original row4: col4 R, col8 R. Shift left 4 => col0,4. Output row4: "R . . . R . . . . . . . ." That's Rs at c

... [4025 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.93)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 245: "A book falls open and its pages splay apart from the spine." k=1 alts: 6/15 4364c1c4

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a uniform background color. Against this background are one or more composite shapes, each consisting of two differently-colored rectangular regions stacked vertically (one color on top, another color on bottom). The two color regions share a horizontal boundary where they meet. Each region may have an irregular outline (not necessarily a perfect rectangle), but the top region and bottom region together form a single connected composite shape. The top region's leftmost column aligns with the bottom region's leftmost column in the input.

Do: For each composite shape in the grid, identify its two color regions: the upper color region and the lower color region. Shift the entire upper color region one column to the left, replacing vacated cells with the background color. Shift the entire lower color region one column to the right, replacing vacated cells with the background color. The two halves spread apart horizontally like opening a hinge — the top half moves left by one cell and the bottom half moves right by one cell. All other cells remain unchanged.

Grid: The grid dimensions remain the same from input to output. The background is a single uniform color filling all cells not occupied by the composite shapes. The grid is large enough that shifting regions by one column does not push any colored cells beyond the grid boundary.

Figurative Descriptions (16 variants)

"A book falls open and its pages splay apart from the spine."

120b: k=1

See: Against a calm, monochrome sky, you see one or more closed books standing upright. Each book has a colored cover on top and a differently colored cover on the bottom, pressed together along a shared spine where the two halves meet horizontally.

Do: A breeze catches each book and it falls open: the top cover fans to the left while the bottom cover fans to the right, each drifting exactly one step away from the spine. The sky fills in the thin gap each half leaves behind. The books breathe open, their two halves parting like wings unfolding from a single crease.

Grid: The sky is a single background color that fills every space not claimed by a book. The world is wide enough that no cover slides off the edge when the books fall open. Everything stays the same size — only the halves shift apart.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10895 chars)

We need to infer rule. From description: books upright with colored cover top and bottom, share spine horizontally. When open, top cover moves left one step, bottom cover moves right one step, leaving gap filled with background color (sky). So input has vertical stacks? Let's examine example.

Input grid 12x12. Colors: B background? In output, many B remain. The "books" likely are vertical rectangles of two colors stacked vertically (top half one color, bottom half another). In input example, we see region of Y (yellow) and R (red) etc.

Let's locate in input example 1: rows 3-6 (0-index?) Actually rows 3-6 have Y region: rows 3-6 (starting at row index 2) columns? Let's map.

I'll index rows 0-11, cols 0-11.

Input rows:
0: B B B B B B B B B B B B
1: B B B B B B B B B B B B
2: B B Y B B B Y Y Y B B B
3: B B Y B B B Y Y Y B B B
4: B B Y Y Y Y Y Y Y B B B
5: B B Y Y Y Y Y Y Y B B B
6: B B R R R R R R R B B B
7: B B R R R R B B R B B B
8-11 all B.

So there are two "books": one with Y top? Actually Y region appears as a shape: top half maybe Y? Let's see shape: rows 2-5 columns 2-8 maybe. The Y region forms a rectangle? rows 2-5, columns 2-8 (7 columns). But note rows 2-3 have Y at col2 and col6-8? Actually row2: col2=Y, col6=Y, col7=Y, col8=Y. Row3 same. Row4: cols2-8 all Y. Row5 same. So shape is like a "book" with top cover maybe Y? But there is also R region rows6-7 columns2-? Row6: cols2-8 all R. Row7: cols2-5 R, col6? row7 col6 = B, col7= B, col8=R. So R region is like bottom cover? The shape is similar but with a missing piece.

In output, we see Y region shifted left? Let's compare output rows 2-5: Output rows2: B Y B B B Y Y Y B B B B (col1? Actually output row2: B Y B B B Y Y Y B B B B). So Y at col1 and col5-7. In input row2 Y at col2 and col6-8. So each Y moved left by 1? Indeed col2->col1, col6->col5, col7->col6, col8->col7. So shift left by 1.

Similarly bottom cover R rows6-7 shift right by 1? Output row6: B B B R R R R R R R B B (R from col2-8 moved to col3-9). Indeed shift right by 1. Row7 output: B B B R R R R B B R B B. Input row7: B B R R R R B B R B B B. Shift right: col2->col3, col3->col4, col4->col5, col5->col6, col6 B stays? Actually B at col6 becomes B at col7? Wait shift right only for R cells, not B. So R at col2-5 become col3-6. R at col8 becomes col9. So matches.

The gap (spine) column originally where the two halves met? In input, the spine is the line where top and bottom halves meet horizontally: they share a row? Actually they are stacked vertically, so spine is the horizontal line between them (row5/6?). The gap after moving: column where they were adjacent now becomes background B. Indeed column2? Let's see after shift, there is a column of B at col2? Output rows2-5 have B at col2 (since Y moved left). Row6-7 have B at col2 (since R moved right). So column2 becomes B (background). So rule: For each vertical stack of two colors (top cover color A, bottom cover color B) that are adjacent ver

... [7937 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 251: "A seed planted in a garden corner grows twin canopies that s ... [18 chars truncated]" k=1 alts: 3/15 4522001f

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: A 3×3 grid with R (red) always at the center cell (row 1, column 1). Three G (green) cells form a 2×2 block together with R, occupying one of the four corner quadrants of the grid. The remaining five cells are black (empty), forming an L-shape along the row and column not occupied by the 2×2 block.

Do: 1. Identify which corner quadrant the 2×2 block (R + 3 G cells) occupies. Note the row index (0 or 2) and column index (0 or 2) of the G cell diagonally opposite R within that block — call these gr and gc. 2. Identify the empty row (the row index, 0 or 2, not shared by the 2×2 block) and the empty column (likewise). 3. Create a 9×9 output grid filled with black. 4. Place a single empty divider: the empty row index × 4 gives the output row that stays black; the empty column index × 4 gives the output column that stays black. 5. The remaining 8 rows and 8 columns form a conceptual 2×2 arrangement of four 4×4 quadrants. Determine the diagonal: if gr and gc are the same (both 0 or both 2), the G diagonal is the main diagonal; if they differ, it is the anti-diagonal. 6. Fill the two 4×4 quadrants along that diagonal entirely with G.

Grid: Input is always 3×3 with values from {., R, G}. Output is always 9×9. The output contains exactly two 4×4 solid green blocks, one entirely empty row, one entirely empty column, and black elsewhere. The empty row and column each occupy one of positions 0, 4, or 8 in the output, corresponding to input positions 0, 1, or 2 scaled by 4.

Figurative Descriptions (16 variants)

"A seed planted in a garden corner grows twin canopies that shade the diagonal."

120b: k=1

See: A small 3×3 garden plot. A red seed sits at the very center, and three green sprouts huddle around it, forming a little square patch in one corner. The rest of the soil is bare — an empty row and an empty column form paths of untilled earth along two edges.

Do: The seed germinates and the garden expands into a vast 9×9 field. The two bare paths from the original plot widen into full irrigation channels — one horizontal, one vertical — that cut across the field, leaving it forever barren where they pass. The channels divide the field into four large planting quadrants. The green sprouts remember which diagonal of the original corner they grew from, and they send their canopy spreading across the two quadrants along that same diagonal, filling each one completely with lush green growth. The other two quadrants remain fallow.

Grid: The small plot is always 3×3 with one red seed, three green sprouts, and bare earth. The grown field is always 9×9. Two irrigation channels (one row, one column) stay barren, and exactly two of the four large quadrants burst into solid green canopy along the diagonal the sprouts originally faced.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (11768 chars)

We need to infer rule. Input 3x3 with R at center? In example1 input:

Row1: . . .
Row2: . R G
Row3: . G G

So R at (2,2) (1-indexed). Greens: positions (2,3), (3,2), (3,3) forming a 2x2 block missing top-left? Actually they form a corner (bottom right) of 2x2 block with R at its top-left? Wait R at (2,2). Greens at (2,3) right of R, (3,2) below R, (3,3) diagonal down-right. So they form an L shape plus diagonal, basically a 2x2 square with R at top-left, greens fill other three cells. That's described as "three green sprouts huddle around it, forming a little square patch in one corner."

Output is 9x9. Let's map.

Output rows 1-9, columns 1-9.

Given output grid (I'll index). Provided:

Row1: . . . . . . . . .
Row2: . G G G G . . . .
Row3: . G G G G . . . .
Row4: . G G G G . . . .
Row5: . G G G G . . . .
Row6: . . . . . G G G G
Row7: . . . . . G G G G
Row8: . . . . . G G G G
Row9: . . . . . G G G G

Interpretation: There is a vertical channel column? The barren (.) columns: column1 all ., column6 all .? Actually column1 all rows are . (yes). Column6 all rows are . (yes). So two channels: column1 and column6 (vertical). Also there is a horizontal channel row1? Row1 all . (yes). Row6? Row6 has . . . . . G G G G, not all . So horizontal channel is row1? Wait description said one horizontal, one vertical. In output, row1 is all . (horizontal channel). Column1 is all . (vertical channel). So they intersect at (1,1) top-left corner.

Quadrants: The 9x9 divided by row1 and column1? Actually row1 and column1 split into quadrants: top-left quadrant is just cell (1,1) which is .; top-right quadrant rows1? Actually row1 is channel, so quadrants are below row1 and right of column1. So quadrants are: upper-left (rows 1? none), upper-right (rows 1? none), lower-left (rows 2-9, col1? column1 is channel, so none), lower-right (rows2-9, cols2-9) is the main area. But output shows green region in two quadrants: one is rows2-5, cols2-5 (a 4x4 block) filled with G. Another is rows6-9, cols6-9 (4x4 block) filled with G. So they are diagonal quadrants relative to the original corner? The original green corner was bottom-right of the 2x2 block. That corresponds to diagonal direction down-right from R. In output, the green fills quadrants that are down-right from the intersection? Actually the intersection is at (row1, col1). The quadrants relative to that are: top-left (none), top-right (rows1, cols2-9) but row1 is channel, so empty. bottom-left (rows2-9, col1) channel. bottom-right (rows2-9, cols2-9) is the main area. But they filled two sub-quadrants within that: rows2-5, cols2-5 (upper-left of bottom-right) and rows6-9, cols6-9 (lower-right). Those are along the main diagonal of the bottom-right quadrant. So they filled the two quadrants along the diagonal direction of original green corner (down-right). So they split the bottom-right quadrant into four 4x4 sub-quadrants (since 8x8 area). They filled the ones that are along the same diagonal as the original green 

... [8777 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.21)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 262: "Each shape is a creature stretching toward a scent on the br ... [40 chars truncated]" k=1 alts: 6/15 470c91de

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains several distinctly colored rectangular shapes on an orange (O) background. Each shape is an almost-complete rectangle — it is missing exactly one corner cell. That missing corner is filled with an azure (A) cell instead of the shape's own color. The azure cell thus sits at one corner of what would be the shape's full bounding rectangle, indicating a direction.

Do: For each colored shape (ignoring orange background and azure markers): 1. Identify the shape's bounding rectangle by finding the minimal rectangle that contains all cells of that color plus the adjacent azure (A) cell that completes the rectangle's missing corner. 2. Determine which corner of the bounding rectangle the azure cell occupies (top-left, top-right, bottom-left, or bottom-right). 3. Complete the shape into a full rectangle by replacing the azure cell with the shape's color. 4. Shift the entire completed rectangle by exactly 1 cell in the direction indicated by the azure corner: top-left corner means shift up 1 and left 1; top-right means shift up 1 and right 1; bottom-left means shift down 1 and left 1; bottom-right means shift down 1 and right 1. 5. Place the shifted rectangle on a blank orange background. Repeat for all shapes. The original positions are cleared to orange.

Grid: The grid dimensions are the same for input and output. The background color is orange (O). All colored shapes remain fully within grid bounds after shifting. Shapes do not overlap.

Figurative Descriptions (16 variants)

"Each shape is a creature stretching toward a scent on the breeze, then lunging one step to catch it."

120b: k=1

See: Against a vast orange desert, colored creatures rest — nearly perfect rectangles, but each one is missing a single corner tooth. In that gap sits a tiny azure firefly, hovering just outside the creature's bite. The firefly marks the direction the creature is yearning toward, like a compass needle the creature can smell but not yet reach.

Do: Each creature swallows the azure firefly, filling in its missing tooth to become a complete rectangle. Now whole, the creature lunges one step diagonally toward where the firefly was — pulled in that direction as if the firefly's glow left a trail. A firefly at the top-left corner pulls the creature up and to the left; one at the bottom-right pulls it down and to the right, and so on. After every creature has lunged, the desert sand (orange) fills in the places they left behind, as if they were never there.

Grid: The desert stays the same size. The sand is always orange. Every creature remains fully within the desert after lunging — none leap off the edge. No two creatures collide.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (12523 chars)

We need to infer rule from example. Let's parse.

We have colors: W maroon, A azure, X grey, O orange desert, B blue, G green, Y yellow, R red.

Metaphor: creatures are rectangles missing a corner tooth (i.e., L-shaped? Actually rectangle missing one corner cell). Azure firefly is a single A cell adjacent to missing corner? The firefly marks direction creature wants to go. Creature fills missing corner (i.e., becomes full rectangle) then moves one step diagonally towards firefly direction (i.e., the direction of the missing corner). After moving, the original rectangle cells become orange (sand). So output shows rectangle shifted diagonally by one cell towards the missing corner, and the original location becomes orange.

Let's examine example 1.

Input grid 10x10.

Let's locate creatures. Look at colors B (blue) forms a 3x3 block? Input lines:

Row0: W W O O O O O O O O
Row1: W A O O O A B B O O
Row2: O O O O O B B B O O
Row3: O O O O O B B B O O
Row4: O O O O O B B B O O
Row5: O O O O O O O O O O
Row6: O X X X X A O O O O
Row7: O X X X X X O O O O
Row8: O X X X X X O O O O
Row9: O X X X X X O O O O

We have two creatures: one blue rectangle missing a corner with azure A at missing corner? Let's see B region: rows2-4 columns5-7 (0-index). That's a 3x3 solid block of B. No missing corner. But there is an A at row1 col5? Actually row1 col5 is A (since row1: W A O O O A B B O O). Wait positions: columns:0 W,1 A,2 O,3 O,4 O,5 A,6 B,7 B,8 O,9 O. So there are two A's: at (1,1) and (1,5). The B block is at rows2-4, cols5-7. The A at (1,5) is directly above the B block, at its top-left corner? The B block missing top-left cell? Actually B block is full 3x3, but there is an A above it, not part of B. Maybe the creature is the B block missing top-left corner (i.e., shape of 3x3 with missing top-left cell) and the A indicates missing corner location. Let's see: If we consider rectangle of size 3x3 that would occupy rows1-3, cols4-6? Not matching.

Let's map coordinates: B block coordinates: rows2-4, cols5-7. That's 3 rows x3 cols. The missing corner would be at (row1,col4) maybe? But A at (1,5) not (1,4). Hmm.

Maybe the creature is the X grey block? X region rows6-9, cols1-5 (a 4x5 rectangle?) Actually rows6-9, cols1-5 are X (grey). That's a 4x5 block (rows6-9 inclusive =4 rows, cols1-5 =5 columns). But there is an A at row6 col5? Row6: O X X X X A O O O O. So A at (6,5) which is rightmost column of X block? The X block missing a corner at (6,5) maybe? The X block is rectangle missing top-right corner? Let's see shape: rows6-9, cols1-5 includes cells (6,1-4) X, (6,5) is A not X. So the rectangle of X would be rows6-9, cols1-5, but missing cell (6,5) which is top-right corner. So the missing corner is at top-right, and the A is there. So creature (grey) will fill missing corner (turn A to X) then move diagonally towards that corner direction: top-right direction means move up and right by one cell (i.e., row-1, col+1). So after transformation, the X block 

... [9547 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 270: "The green jewel is a hall of mirrors, and the red dancer see ... [32 chars truncated]" k=2 alts: 12/15 4938f0c2

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains exactly two objects on a black background: (1) a green 2×2 block, and (2) a red shape made of connected red cells. The red shape is adjacent to the green block, positioned on one side (or in one quadrant) relative to the green center. The green block acts as a central pivot point.

Do: Find the center of the green 2×2 block (the point where its four cells meet). Reflect the entire red shape to produce four-fold symmetry around this center point: (1) mirror the red shape horizontally across the vertical axis through the green center, (2) mirror the red shape vertically across the horizontal axis through the green center, and (3) mirror it diagonally (equivalently, apply both horizontal and vertical reflections). For each original red cell at position (r, c), compute the three reflected positions using the green center (cr, cc): horizontally reflected to (r, 2·cc − c − 1), vertically reflected to (2·cr − r − 1, c), and doubly reflected to (2·cr − r − 1, 2·cc − c − 1). Place red cells at all reflected positions. Keep the green block unchanged. If a reflected copy already exists (the input already has symmetry in some direction), the output is unchanged in that region.

Grid: The grid dimensions match the input exactly (variable size). The background is black (.). The green 2×2 block remains in place. The output grid contains the original red shape plus up to three reflected copies, forming four-fold (bilateral + rotational) symmetry around the green block's center.

Figurative Descriptions (16 variants)

"The green jewel is a hall of mirrors, and the red dancer sees her reflections on every wall."

120b: k=2

See: A red dancer strikes a pose near a small green jewel sitting on a dark stage. The jewel has four perfectly polished faces — it is a tiny mirror box at the heart of the world.

Do: The green jewel awakens and casts the dancer's image outward in every direction it can reflect — left and right, up and down, and corner to corner. Wherever the dancer stands, her reflections appear on the opposite sides of the jewel, as if she is performing inside a kaleidoscope. The jewel itself never moves — it only watches and reflects. If a reflection would land where the dancer (or another reflection) already stands, nothing changes — mirrors don't argue over who arrived first.

Grid: The dark stage stays the same size. The green jewel remains fixed at its position, glowing quietly. The red dancer now appears to be surrounded by her own reflections, forming a symmetrical constellation that honors the jewel as its center. Every quadrant around the jewel tells the same story, just viewed from a different mirror.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (9235 chars)

We need to infer rule: There's a green jewel (G) of size 2x2 (as in examples). The red dancer (R) pattern is reflected across the jewel in all 8 directions (horizontal, vertical, diagonal) producing symmetric pattern around the jewel. The jewel stays same. The reflections are placed such that for each red cell, its mirror across the center of the jewel (the 2x2 block) is placed. The jewel is at some coordinates. In example1, input grid 10x10. Jewel at rows 3-4? Let's index 0-based. Input example1 rows: row0 ". R R . . . . . . ." etc. Jewel appears rows 3 and 4 (0-index) columns 4 and5? Actually row3: ". . . . G G . . . ." So columns 4,5 are G. Row4 same. So jewel at (3,4)-(4,5). Output shows reflections of R pattern around jewel: original R cluster near top left of jewel? Let's see input R positions: rows0-2 columns1-3 etc. Output adds R's mirrored to bottom right etc. Indeed they reflect across both axes.

In example2, larger grid 20x30? Jewel at rows7-8 columns9-10 (0-index). Input R pattern is a plus shape? Actually there is a vertical line of R at columns5 rows3-6, plus a horizontal line at row6 columns5-8, making a T shape? Output shows reflections across jewel both horizontally and vertically, creating pattern repeated in four quadrants.

Thus rule: For each red cell, compute its offset relative to the center of the jewel (the center point between the four G cells). Then place a red cell at the symmetric position across the jewel: i.e., reflect the offset sign (dx, dy) to opposite signs, generating up to 4 copies? Actually they reflect in all directions: original stays, plus reflections across vertical axis (mirror left-right), across horizontal axis (up-down), and across both (diagonal). So total up to 4 copies (including original) making symmetric about the jewel's center.

Implementation: For each red cell at (r,c), compute its vector to the jewel center. Jewel center is between the four cells: coordinates (r0+0.5, c0+0.5) where top-left G at (r0,c0). Then reflected positions are (center - (dx,dy))? Actually original offset = (r - center_y, c - center_x). Mirror across vertical axis: (r - center_y, -(c - center_x)). Mirror across horizontal: (-(r - center_y), c - center_x). Mirror across both: (-(r - center_y), -(c - center_x)). Then round to integer positions? Since center is half-integer, offsets are half-integer, resulting positions integer. Works.

Now apply to test input.

Test grid size 18 rows x14 columns? Actually lines show 18 rows, each 14 columns. Jewel: rows6-7 columns6-7 are G (0-index). Let's index.

Rows (0-based):
0: all .
1: all .
2: ". . R R R R . . . . . . . ."
3: ". . . R . . . . . . . . . ."
4: ". . . . R . . . . . . . . ."
5: ". . . . R R . . . . . . . ."
6: ". . . . . . G G . . . . . ."
7: same as row6
8-17 all .

Now list R coordinates.

Row2: columns2,3,4,5 are R (since ". . R R R R . . ...") So (2,2),(2,3),(2,4),(2,5)
Row3: column3 is R (". . . R ...") => (3,3)
Row4: column4 is R => (4,4)
Row5: columns4,5 

... [6298 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 280: "A compass rose casts reflections of a traveler into every ca ... [14 chars truncated]" k=1 alts: 0/15 4c5c2cf0

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains two distinct colored objects on a black background. The first is an X-shaped directional marker: four cells of one color arranged at the four diagonal neighbor positions around an implicit center cell (which is black). This forms a small diamond/X pattern spanning a 3x3 area. The second object is an asymmetric shape of a different color, located in one of the four diagonal quadrants relative to the X-marker's center. The X-marker and the shape do not overlap.

Do: 1. Locate the X-shaped marker: find four same-colored cells arranged at positions (r-1,c-1), (r-1,c+1), (r+1,c-1), (r+1,c+1) for some center (r,c). This center (r,c) is the symmetry anchor. 2. Identify the other colored shape. Compute each of its cell positions as offsets (dr, dc) relative to the center (r,c). 3. Reflect the shape into all four diagonal quadrants by generating four copies of each offset: (dr, dc), (-dr, dc), (dr, -dc), and (-dr, -dc). 4. Place all four reflected copies of the shape onto the grid in addition to the original X-marker, which remains unchanged. The background stays black everywhere else.

Grid: The grid dimensions match the input exactly. The background is black (.). The X-shaped marker remains at its original position in the output. The grid must be large enough to contain all four reflections of the shape; reflections that would fall outside the grid boundary do not occur in valid inputs.

Figurative Descriptions (16 variants)

"A compass rose casts reflections of a traveler into every cardinal shadow."

120b: k=1

See: On a dark plain, a compass rose glows—four sentinel gems set at the diagonal points around an invisible hub, forming a small X of light. Somewhere in one quadrant of the plain, a wandering figure (a different-colored shape) has settled, as if a traveler who arrived from just one direction.

Do: The compass rose is a hall of mirrors aimed at the four diagonal winds. When the traveler stands in one quadrant, the rose catches their silhouette and projects it into the three empty quadrants—flipping it across the rose's invisible crosshairs so that every quadrant receives its own reflected twin. The traveler multiplies into four, each a mirror image appropriate to its quadrant, while the compass rose itself remains unmoved and unchanged at the center of it all.

Grid: The dark plain keeps its original dimensions. The compass rose stays planted where it was. Four versions of the traveler now populate the grid, arranged symmetrically around the rose's hub like four dancers reflected in a pond. The darkness fills everything the light does not touch.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10773 chars)

We need to understand the transformation. In example 1, input grid shows a pattern: there is a central "compass rose" of four sentinel gems at diagonal points around an invisible hub, forming a small X of light. In example, the sentinel gems are maybe color A? Actually in input example 1, we see A's forming a small X? Let's examine.

Input example 1 grid (14x14). Let's locate A's: At row 3 (0-index?), row index 3 (fourth row) column 8 has A. Row 4: columns 7,8,9 have A A A. Row5: columns 6,7,8 have A A A. That's like a small shape maybe a triangle? Actually looks like a 3x3 block offset? Let's map coordinates.

Rows 0-13.

Row3 col8 A.
Row4 col7-9 A A A.
Row5 col6-8 A A A.

That's like a diagonal line of A's forming a shape maybe a "compass rose"? The description: four sentinel gems set at diagonal points around an invisible hub, forming a small X of light. So sentinel gems are at four diagonal points around a center. In example, maybe the A's are the compass rose? Let's see pattern: The A's form an X shape? Let's list coordinates: (3,8), (4,7),(4,8),(4,9), (5,6),(5,7),(5,8). That's not symmetric X; it's like a small shape maybe a rotated L? Hmm.

But also there are G's forming a shape elsewhere: rows 6-8 have G pattern: row6 col4 G, col6 G? Actually row6: ". . . . G . G . . . . . . . ." So G at col4 and col6. Row7: ". . . . . G . . . . . . . . ." G at col5. Row8: same as row6. That's a small X shape of G's centered at (7,5). That's likely the compass rose (four sentinel gems). Indeed G's are at diagonal points around center (7,5). The pattern: positions (6,4), (6,6), (7,5), (8,4), (8,6). That's a plus? Actually that's a small X: top-left, top-right, center, bottom-left, bottom-right. That's 5 points forming an X shape. So sentinel gems are G's.

Thus the compass rose is G's. The traveler is a different-colored shape located in one quadrant relative to the rose. In input, there is a shape of A's located in upper right quadrant relative to the rose? Let's find rose center: (7,5). Quadrants: upper-left rows<7, cols<5; upper-right rows<7, cols>5; lower-left rows>7, cols<5; lower-right rows>7, cols>5.

A shape coordinates: rows 3-5, cols 6-9. Those are rows less than 7, cols greater than 5 => upper-right quadrant. So traveler is A shape.

In output, we see A shape replicated in all four quadrants, mirrored appropriately.

Let's verify output: A's appear also in other quadrants: In output rows 3-5 same as input but shifted? Actually output row3: ". . . . . . . . A . . . . . ." Wait row3 col8? Let's check output row3: ". . . . . . . . A . . . . . .". That's A at col8 (same as input). Row4: ". A A A . . . A A A . . . .". So now there are two groups of A's: one at left side (col1-3?) Actually row4: positions col1,2,3 are A A A (that's left side). And also col7-9 A A A (right side). So left side group is mirrored to upper-left quadrant. Row5: ". . A A A . A A A . . . .". So left group at col2-4? Actually row5 col2,3,4 A A A (mir

... [7938 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.42)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 290: "Two armies join forces under a shared banner" k=1 alts: 8/15 506d28a5

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid consists of two equally-sized rectangular grids stacked vertically, separated by a single horizontal row of yellow (Y) cells that acts as a divider. The top grid contains red (R) cells and black (.) cells. The bottom grid contains blue (B) cells and black (.) cells.

Do: Remove the yellow divider row to isolate the top grid and bottom grid. Create an output grid of the same dimensions as either half. For each cell position, if the top grid has red (R) at that position OR the bottom grid has blue (B) at that position (or both), set the output cell to green (G). If neither grid has a colored cell at that position (both are black), set the output cell to black (.). This is a logical OR operation across the two grids.

Grid: The input grid has dimensions (2n+1) rows by m columns, where the middle row is entirely yellow. The output grid has dimensions n rows by m columns. In the examples, n=4 and m=5, producing a 9×5 input and a 4×5 output.

Figurative Descriptions (16 variants)

"Two armies join forces under a shared banner"

120b: k=1

See: A golden river runs across the middle of the battlefield, dividing two kingdoms. Above the river, the Red Army has pitched its tents on certain squares. Below the river, the Blue Army holds its own positions. The rest of the land lies empty and dark.

Do: The war is over — the golden river dries up, and the two kingdoms merge into one. Wherever either army had claimed ground, the united banner of green now flies. Only the squares that neither army ever touched remain barren darkness. Every soldier, red or blue, becomes green in the new alliance.

Grid: The golden river is a single row splitting the world in half. The united kingdom is exactly the size of one half. Every plot of land answers one question: did anyone fight here? If so, it blooms green.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (5173 chars)

We need to infer rule. Description: golden river is a single row splitting world in half. Input grid seems to have a row of Y's (yellow) representing river. Above river there are red R positions (army). Below river there are blue B positions. Output is a grid of size equal to one half (either above or below?) The description: "The united kingdom is exactly the size of one half." So output grid size equals half the input rows (excluding river?). In example 1, input grid dimensions? Let's count rows.

Example 1 Input lines:
". . . . ."
". R R R ."
". R . . ."
"R R R . R"
"Y Y Y Y Y"
"B B . . B"
"B B . B B"
"B . B . B"
". B . B B"

That's 9 rows, each 5 columns. River row is row 5 (Y's). So above river rows = 4 rows, below river rows = 4 rows. Output grid is 4 rows, 5 columns. Indeed output given 4 rows.

Output rows:
"G G . . G"
"G G G G G"
"G G G . G"
"G G G G G"

So output corresponds to merging armies: any cell that had either red (above) or blue (below) becomes green. Cells that were empty (.) remain .? But output shows many G's where there were reds or blues. Let's verify mapping.

Take position (row1,col1) input top half row1 col1 = '.' (black). Output row1 col1 = G. Wait that's not matching. Let's map coordinates: Output row1 corresponds to input row1? Actually output row1 corresponds to top half row1? Let's compare.

Input top half rows (rows1-4):
1: . . . . .
2: . R R R .
3: . R . . .
4: R R R . R

Output rows:
1: G G . . G
2: G G G G G
3: G G G . G
4: G G G G G

Now compare each cell: For cell (1,1) input '.' becomes G. So rule not just red->green. Maybe any cell that is not yellow (river) becomes green regardless of being empty? But output has some '.' still? Actually output row1 col3 is '.' (input row1 col3 '.'), output row1 col4 '.' (input '.'), others G. So not all become G.

Let's see pattern: maybe output is union of red positions (above) and blue positions (below) projected onto same half? Since output size equals half, maybe we overlay the two halves onto each other (like combine). For each column, we take red from above and blue from below, and if either has a soldier, output green. If both empty, output '.'.

But need to align rows: top half rows correspond to same rows as bottom half? Probably we overlay row i from top with row i from bottom (mirrored?) Let's test.

Bottom half rows (rows6-9):
6: B B . . B
7: B B . B B
8: B . B . B
9: . B . B B

Now overlay row1 (top) with row6 (bottom) maybe? Let's compute union: For each column, if either top row1 colj is R (or any non-empty) or bottom row6 colj is B (or any non-empty) then output G else '.'.

Row1 top: . . . . .
Row6 bottom: B B . . B
Union: columns 1,2,5 have B => G. columns3,4 empty => . So output row1 would be G G . . G. That matches output row1! Good.

Row2 top: . R R R .
Row7 bottom: B B . B B
Union: col1 top '.' bottom B => G. col2 top R bottom B => G. col3 top R bottom . => G. col4 top R bottom B => G. col5 top . bottom B => G. So all G => r

... [2305 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.13)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 292: "A ball bounces off a wall like light reflecting off a mirror ... [1 chars truncated]" k=3 alts: 8/15 508bd3b6

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains three elements: (1) a black (.) background, (2) a solid rectangular red (R) wall along one edge of the grid (it can be along the bottom, right, left, or top — occupying full rows or full columns), and (3) exactly two or three azure (A) cells arranged in a diagonal line. The A cells define a trajectory direction — moving from the first A to the last A gives a diagonal direction vector (e.g., down-right, down-left, up-right, or up-left). The A diagonal is positioned in the open black region, with the red wall nearby on one side.

Do: 1. Identify the diagonal direction of the A cells by looking at consecutive A positions. Compute the direction vector (row_delta, col_delta), each being +1 or -1. 2. Identify the red wall: it is a solid rectangular block of R cells occupying entire rows or columns along one edge of the grid. Determine which edge it lies on (top, bottom, left, or right). 3. Starting from the last A cell in the direction of travel, extend the trajectory one step at a time in the same diagonal direction, placing a green (G) cell at each new position. Continue until the next step would enter the red wall region or exit the grid. 4. When the trajectory reaches the red wall, reflect the ball: reverse the component of the direction vector that is perpendicular to the wall. If the wall is horizontal (top or bottom edge), negate the row_delta; if the wall is vertical (left or right edge), negate the col_delta. 5. Continue placing G cells in the new reflected direction, one step at a time, until the trajectory exits the grid boundary. 6. Leave all original A cells and R cells unchanged. Only add G cells on previously black cells.

Grid: The grid is rectangular (not necessarily square), typically around 12x12. The background is black (.). The red wall is a contiguous rectangular block spanning the full width or full height of the grid along one edge, and can be multiple rows or columns thick. The A cells and G cells occupy the open (non-red) portion of the grid. The output grid has the same dimensions as the input.

Figurative Descriptions (16 variants)

"A ball bounces off a wall like light reflecting off a mirror."

120b: k=3

See: A black night sky stretches across the grid. Along one edge, a solid red cliff face stands immovable — a wall of crimson stone. Out in the open darkness, a few azure sparks trace a diagonal path, like a ball mid-flight, hurtling toward the cliff.

Do: The azure sparks show where the ball has already been — its past trajectory frozen in place. Follow the ball's momentum forward, leaving a green phosphorescent trail as it streaks through the darkness. When the ball reaches the red cliff, it bounces — ricocheting like a billiard ball off a cushion, its angle of incidence equal to its angle of reflection. The green trail continues in the new direction until the ball flies off into the void beyond the grid's edge. The cliff and the original sparks remain untouched.

Grid: The arena is a rectangular court draped in black. The red cliff is a thick, immovable barrier running along one entire side. The azure marks and the green trail live only in the open darkness beyond the cliff. The court's dimensions stay fixed.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (13382 chars)

We need to infer rule. From description: black background, red cliff along one side (full side). Azure (A) marks past trajectory of ball. Need to extend ball's path forward, leaving green (G) trail, bounce off red wall with equal angle, continue until off grid. The red wall and original A remain unchanged.

Thus we need to detect direction of ball from A positions (a line of A's). They represent previous positions. Determine direction vector (dx, dy) from consecutive A's. Then extend forward from last A (the tip) in same direction, marking G until hitting red wall (R). At wall, reflect direction (dx, dy) -> reflect across wall normal. Since wall is along one side (full side). Which side? In examples, red wall is at bottom rows (example1 rows 10-12). In example1, A's are diagonal from (row3,col1) etc? Let's map coordinates: Example1 input grid 12 rows, 12 columns. Red rows at rows 10-12 all columns. A's at (row3,col1), (row4,col2), (row5,col3). That's a diagonal down-right. So direction vector (1,1). The ball moves forward from last A at (5,3) direction (1,1) towards red wall at bottom. It will hit red wall at row10? Let's compute: positions: (6,4) G, (7,5) G, (8,6) G, (9,7) G, (10,8) would be red wall cell (since row10 all R). At that point bounce. Reflection off horizontal wall (bottom) would invert dy (vertical component) while keep dx same? Actually wall is horizontal line (row10). Incidence angle equal reflection: direction (1,1) hitting horizontal wall => reflect vertical component: (1,-1). So after bounce, continue with (1,-1) moving up-right. Starting from cell just before wall? In output, we see G's placed at (6,4), (7,5), (8,6), (9,7) as above, and also after bounce they have G at (9,8)? Wait output shows G at (9,8?) Let's examine output rows: Row6 (index6) has G at col4? Actually row6 line: ". . . G . . . . . G . ." That's row6 (original row6). G at col4 and col10? Wait there are two G's per row after bounce? Let's map.

Let's write coordinates (row, col) 1-indexed.

Output rows:

Row1 all .
Row2 all .
Row3 A at col1 unchanged.
Row4 A at col2, G at col4? Actually row4: ". A . . . . . . . . G". So G at col12? Wait row4 length 12. Row4 shows G at last column (col12). That's after bounce maybe.

Row5: ". . A . . . . . . . G ." G at col11.

Row6: ". . . G . . . . . G . ." G at col4 and col10.

Row7: ". . . . G . . . G . . ." G at col5 and col9.

Row8: ". . . . . G . G . . . ." G at col6 and col8.

Row9: ". . . . . . G . . . . ." G at col7.

Rows10-12 red.

Observations: The path after bounce seems to go up-left? Actually after hitting bottom, direction becomes ( -1, 1 )? Let's compute: Starting from last A at (5,3). Continue (1,1): (6,4) G, (7,5) G, (8,6) G, (9,7) G, then bounce at row10 (R). After bounce, direction becomes (-1,1) (up-right). So next positions: (9,8) G, (8,9) G, (7,10) G, (6,11) G, (5,12) G, then off grid. Indeed output shows G at (9,8) row9 col8? Row9 has G at col7 only, not col8. Wait row9 line: ". . . . . . G . . .

... [10391 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.50)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 319: "Russian nesting dolls are packed away, each one's importance ... [24 chars truncated]" k=1 alts: 15/15 5587a8d0

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has an orange (O) background containing several distinct non-orange colored shapes, each made up of a different number of cells. Each color appears exactly once as a distinct group of cells scattered across the grid. The key feature of each shape is simply how many cells it occupies — its position and arrangement in the input do not matter.

Do: 1. Identify all distinct non-orange colors in the input and count the number of cells each color occupies. 2. Rank the colors from most cells (largest) to fewest cells (smallest). 3. Let n be the number of distinct colors. The output grid is (2n−1) × (2n−1). 4. Build the output as concentric rectangular rings, each one cell wide, nesting inward: the largest color forms the outermost ring, the second-largest forms the next ring inward, and so on, with the smallest color filling the single center cell. 5. If two colors have the same cell count, their relative order may be ambiguous; otherwise, strict descending cell count determines ring order from outside to inside.

Grid: The output grid is always a square of side length (2n−1), where n is the number of distinct non-orange colors in the input. There is no background color in the output — every cell belongs to one of the concentric rings. The grid is symmetric about both axes and both diagonals.

Figurative Descriptions (16 variants)

"Russian nesting dolls are packed away, each one's importance measured by its weight."

120b: k=1

See: Scattered across an orange workshop table lie the disassembled pieces of a set of nesting dolls — each painted a different color, some larger and heavier than others. The dolls have tumbled out of order; their sizes are all you need to know, not where they landed.

Do: The dollmaker gathers the pieces and reassembles the set. The heaviest doll becomes the outermost shell, the next heaviest nestles inside it, and so on, each one cradled snugly within the last, until the tiniest doll — light as a single breath — sits alone at the very heart.

Grid: The finished nesting doll, viewed from above, is a perfect square window into concentric colored shells. Each shell is exactly one layer thick, and the whole assembly grows symmetrically outward from the tiny center doll. The workshop table is gone — every inch of the view belongs to a doll.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (9338 chars)

We need to infer rule. Input grid contains various colors; output is a small square of concentric layers based on "heaviest" meaning maybe frequency? In example1: Input 6x6 with colors O, Y, M. Output 3x3 with Y outer, M center? Actually output shows Y border, M center. So they identified heaviest doll = outermost shell = most frequent color? Let's count frequencies in example1 input.

Let's compute: Grid 6x6.

Rows:
1 O O O O O O -> 6 O
2 O O O Y Y O -> O4? actually O O O Y Y O => O4? positions: O O O (3) Y Y O (1 O) => O4, Y2
3 O O O Y Y O => same O4, Y2
4 M O O O O O => M1, O5
5 M M O O O O => M2, O4
6 O O O O O O => O6

Total O: row1 6 + row2 4 + row3 4 + row4 5 + row5 4 + row6 6 = 29? Let's sum: 6+4=10, +4=14, +5=19, +4=23, +6=29. Y: row2 2 + row3 2 =4. M: row4 1 + row5 2 =3.

Heaviest (most frequent) is O (29). But output outermost is Y, not O. So maybe heaviest means "largest size" defined by something else like area of contiguous region? Let's examine shapes. Input likely contains clusters of colors representing pieces of dolls. The "size" maybe number of cells in each connected component (4-neighbor?). Let's find connected components for each color.

We need to see clusters: O appears everywhere, likely background. Y appears as a 2x2 block at positions (2,4-5) and (3,4-5). That's a 2x2 square. M appears as two cells at (4,1) and (5,1-2). That's an L shape of 3 cells. So Y region size 4, M region size 3. O region is huge but maybe considered background not a doll? The description: pieces of nesting dolls each painted a different color, some larger and heavier than others. The heaviest doll becomes outermost shell. So we need to sort pieces by size (area) descending, then create concentric squares with those colors, smallest at center.

In example, Y size 4 > M size 3, so Y outer, M inner. Output is 3x3 with Y border and M center. That matches.

Thus rule: Identify distinct non-background colors? But which is background? Possibly the color that appears most (O) is the table, to be ignored. The remaining colors are pieces. Sort by component size (number of cells). Then produce a square of size (2* n -1?) Actually with two colors we got 3x3. For k colors, output size = 2*k -1? Let's test: k=2 => size 3. Yes.

If there were 3 colors, size would be 5, etc. The output is concentric layers each one cell thick, outermost color = largest piece, then next, etc, innermost = smallest piece.

Now test input: we need to identify colors other than background. Determine background color: likely the most frequent color overall. Let's count frequencies.

We have colors: O, W, R, X, A, B, . (black), maybe others.

Let's list grid 8x8.

I'll index rows.

Row1: O O W O O R R R
Row2: O O W O O R O R
Row3: X O O O O O O O
Row4: X X O A O O O O
Row5: X O O O O . O O
Row6: O O B O O . . O
Row7: O B B B O O . .
Row8: B B B B B O O .

Let's count each color.

Initialize counts.

We'll go row by row.

Row1: O x? positions: O,O,W,O,O,R,R

... [6590 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.07)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 322: "A moth is drawn to a flame, leaving a shadow on the wall whe ... [20 chars truncated]" k=1 alts: 3/15 56dc2b01

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains three key elements on a black background: (1) a red line (R) — either horizontal or vertical — that spans the full width or height of the grid, acting as a fixed wall; (2) a green shape (G) — an irregular connected figure made of green cells; and (3) empty black space between and around them. The green shape is separated from the red line by one or more rows/columns of empty space, positioned somewhere on the opposite side of the red line from the nearest edge.

Do: 1. Identify the red line and its orientation (horizontal or vertical). 2. Identify the green shape and note its bounding box and internal pattern. 3. Slide the entire green shape along the axis perpendicular to the red line, moving it toward the red line, until the green shape is directly adjacent to the red line (no gap remains between them). 4. Remove the green shape from its original position (replace with black). 5. Place a solid azure (A) line on the far side of the green shape — that is, on the edge of the green shape's bounding box opposite the red line. The azure line runs parallel to the red line and spans exactly the same length as the red line (full row or full column width). The azure line occupies the single row or column immediately after the green shape's bounding box, on the side away from the red line. 6. All other cells remain unchanged.

Grid: The grid dimensions vary across instances (can be rectangular in either orientation). The background is black (.). The red line always spans the full width of the grid (if horizontal) or the full height (if vertical), occupying exactly one row or column. The green shape fits entirely on one side of the red line. The output grid has the same dimensions as the input grid.

Figurative Descriptions (16 variants)

"A moth is drawn to a flame, leaving a shadow on the wall where it finally rests."

120b: k=1

See: In the darkness, a red ember glows — a bright, unbroken line of heat stretching from edge to edge. Somewhere in the void, a green moth flutters, its irregular wings catching no light. Between the moth and the ember lies empty night.

Do: The moth cannot resist the warmth. It drifts through the darkness toward the red ember, crossing all the empty space until it presses its body right against the glowing line. Where the moth once hovered, only darkness remains — no trace of its departure. But the journey leaves a mark: an azure shadow crystallizes along the moth's trailing edge, a cool afterimage the exact length of the ember, as if the night itself remembers the last place the moth paused before surrendering to the heat.

Grid: The void (black) can stretch in any direction. The ember (red) is an unbroken line of fire spanning the full breadth of the world. The moth (green) is an irregular living shape that always begins on one side of the ember. The shadow (azure) is born only after the moth completes its journey.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10944 chars)

We need to infer rule from example. Input grid is 16 rows? Let's count. Example input shows many rows: appears 16 rows? Actually list: rows from top: first row ". . . . ." (5 columns). Then many rows until row with "R R R R R". So grid height maybe 16? Let's count rows: I'll enumerate:

1 . . . . .
2 G G . . .
3 G . . . .
4 G G . G G
5 . G G G .
6 . . . . .
7 . . . . .
8 . . . . .
9 . . . . .
10 . . . . .
11 . . . . .
12 . . . . .
13 . . . . .
14 . . . . .
15 . . . . .
16 R R R R R
17 . . . . .

Actually there are 17 rows? Wait after row 16 with R line, there is row 17 ". . . . .". So total 17 rows, 5 columns.

Output grid: also 17 rows, 5 columns. Let's list output rows:

1 . . . . .
2 . . . . .
3 . . . . .
4 . . . . .
5 . . . . .
6 . . . . .
7 . . . . .
8 . . . . .
9 . . . . .
10 . . . . .
11 A A A A A
12 G G . . .
13 G . . . .
14 G G . G G
15 . G G G .
16 R R R R R
17 . . . . .

So transformation: The red line (R) is at row 16 originally, spanning full width. In output, red line stays same row 16 unchanged. The green shape originally appears in rows 2-5 (some shape). In output, the green shape moved down to rows 12-15? Actually output rows 12-15 contain the same pattern as original rows 2-5. Indeed row 12: G G . . . matches original row2. Row13: G . . . . matches original row3. Row14: G G . G G matches original row4. Row15: . G G G . matches original row5. So the green shape moved down to just above the red line, with one row of A (azure) inserted between them? Actually there is row 11 with A A A A A (azure line) exactly same length as red line, placed directly above the green shape? Wait order: In output, from top: rows 1-10 empty, row11 azure line, rows12-15 green shape, row16 red line, row17 empty.

In input, green shape was at top (rows2-5), red line at row16, empty rows elsewhere. So transformation: Move green shape down until it is directly above red line, leaving exactly one row of azure (A) between green shape and red line? Actually azure line appears above green shape, not between green and red. Let's examine description: Moth (green) moves towards ember (red) crossing empty space until it presses its body right against the glowing line. Where the moth once hovered, only darkness remains. But the journey leaves a mark: an azure shadow crystallizes along the moth's trailing edge, a cool afterimage the exact length of the ember, as if the night itself remembers the last place the moth paused before surrendering to the heat.

Interpretation: The green shape moves downwards (assuming ember is horizontal line). It ends up adjacent to red line (touching). The original location of green becomes empty (black). The azure line appears at the previous location of the green shape's trailing edge? Actually "azure shadow crystallizes along the moth's trailing edge, a cool afterimage the exact length of the ember". The trailing edge is the side opposite direction of movement? If moving downwards, trailing edge is top side of green shape. So azure li

... [7974 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.03)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 349: "A heartbeat grows louder, peaks, then fades — and the ... [37 chars truncated]" k=3 alts: 11/15 5d588b4d

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input is a grid where the top row contains some number N of colored cells (all the same color) packed to the left, followed by black cells. All remaining rows below are entirely black. The grid has width W and a variable number of rows.

Do: 1. Count N, the number of colored cells in the top row. Note the color C and the grid width W. 2. Generate a linear sequence of cells by producing runs of C-colored cells with lengths 1, 2, 3, ..., N, ..., 3, 2, 1 (ascending from 1 to N, then descending back to 1). Separate each consecutive pair of runs with exactly one black cell. 3. If the total length of this sequence is not a multiple of W, pad the end with black cells until it is. 4. Wrap this linear sequence into rows of width W. The resulting rows form the output grid.

Grid: The output grid has the same width W as the input. The number of output rows equals the total padded sequence length divided by W. The background color is black (.). Only one non-black color appears, matching the color from the input's first row.

Figurative Descriptions (16 variants)

"A heartbeat grows louder, peaks, then fades — and the EKG paper scrolls it across the page."

120b: k=3

See: A row of colored pulses sits at the top of a dark monitor, like an initial charge of energy. Below stretches silence — a flatline of black.

Do: The pulse begins to beat. It starts faint — a single colored blip — then each successive beat grows one pulse stronger: tap, tap-tap, tap-tap-tap, building up to the full strength shown in that first row. At its peak the heartbeat holds, then winds back down symmetrically — each beat losing one pulse until it fades to a single blip again. Between every heartbeat, one moment of silence (a black cell) separates the rhythm. This entire cardiac rhythm is recorded onto the EKG tape.

Grid: The EKG tape is exactly as wide as the monitor. The rhythm is printed left to right, and when the tape reaches the right edge it scrolls to the next line. If the last line has leftover space after the final heartbeat fades, the tape fills with flatline silence to the edge.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (11396 chars)

We need to understand the transformation rule from examples.

We have input grid with top row containing a sequence of colored cells (same color) then blanks (.) rest of row. Rest rows are empty (all .). Output shows a pattern of that color arranged in a "heartbeat" pattern across rows, like a triangular wave.

Let's analyze Example1.

Input top row: G G G G G G . . . . . (6 G then 5 .) width 11? Actually count: "G G G G G G . . . . ." that's 11 columns. Input grid 9 rows? Actually given 9 rows of . after top row. Output grid appears 5 rows? Let's count output rows: they show 5 rows. Output rows:

Row0: G . G G . G G G . G G
Row1: G G . G G G G G . G G
Row2: G G G G . G G G G G .
Row3: G G G G . G G G . G G
Row4: . G . . . . . . . . .

So output width same as input width (11). Height seems 5 rows (maybe number of beats?). The pattern: looks like a "pulse" moving downwards? Let's map positions.

Let's index columns 0-10.

Row0: G . G G . G G G . G G
Positions with G: 0,2,3,5,6,7,9,10 (8 cells). Row1: G G . G G G G G . G G => positions 0,1,3,4,5,6,7,9,10 (9 cells). Row2: G G G G . G G G G G . => positions 0,1,2,3,5,6,7,8,9 (9 cells). Row3: G G G G . G G G . G G => positions 0,1,2,3,5,6,7,9,10 (9 cells). Row4: . G . . . . . . . . . => position 1 only.

Seems like a pattern of a "heartbeat" shape: start with single G at column 1? Actually bottom row row4 has a single G at column1 (maybe the final single blip). The top row row0 has pattern starting with G at column0 then blank then two G etc.

Let's think of the metaphor: top row is initial pulse row (full strength). The rhythm is printed left to right, with each beat consisting of increasing number of pulses (colored cells) from 1 up to full length (the number of colored cells in top row?), then decreasing back to 1, with a black cell separating beats. The output grid is like a raster of this sequence across rows, wrapping to next line when reaching edge.

Essentially we need to generate a 1D sequence of colors and blanks (black) representing the heartbeat pattern, then fill into grid rows left to right, top to bottom, with width equal to input width. The input provides the "full strength" pattern: a contiguous block of colored cells (maybe length N) at the start of the top row, rest blanks. That defines the color and the length of the full pulse (max number of colored cells in a beat). The beats are built by starting with 1 colored cell, then 2, then 3, ... up to N, then N-1,... down to 1. Between each beat, there is a single black cell (.) as separator. The pattern repeats? The description says "the entire cardiac rhythm is recorded onto the EKG tape." Probably just one full cycle (up and down) maybe? Let's test with Example1: N = 6 (since 6 G). So beats lengths: 1,2,3,4,5,6,5,4,3,2,1. That's 11 beats. Between each beat, a black cell separator, so total cells = sum of lengths + (number of beats -1) separators? Actually separators between beats, not after last? So total cells = (1+2+...+

... [8444 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.01)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 363: "A herald's banner unfurls across a great hall, its emblem re ... [38 chars truncated]" k=1 alts: 15/15 62b74c02

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a rectangular pattern of colored cells on the left side, followed by a region of black (.) cells extending to the right. The pattern has a key property: in each row, the first and last cells of the pattern share the same color, forming a 'border' color for that row. The interior cells between these matching border cells may be different colors.

Do: 1. Identify the pattern block on the left side of the grid (all columns before the black/dot region). Note its width W and the total grid width T. 2. For each row, identify the border color, which is the color of the first (and last) cell of the pattern block. 3. Construct the output row as follows: place the original pattern block at the left edge, place a mirrored copy of the pattern block at the right edge (so columns T-W through T-1 are the same as columns 0 through W-1), and fill all cells in between with that row's border color. 4. Repeat for every row.

Grid: The output grid has the same dimensions as the input grid (same number of rows and columns). The height equals the height of the input pattern. The width equals the full grid width including the dot region. No black cells remain in the output; they are all replaced by either the repeated border color or the mirrored pattern.

Figurative Descriptions (16 variants)

"A herald's banner unfurls across a great hall, its emblem reflected in the mirror at the far wall."

120b: k=1

See: On the left edge of the hall hangs a narrow heraldic banner — each horizontal stripe bears a distinct coat of arms, framed on both sides by that stripe's own house color. The rest of the hall stretches away into darkness, an empty corridor of black stone.

Do: The banner is carried in procession across the hall. As it moves, each stripe leaves a ribbon of its house color trailing behind it, like a carpet unrolling. When the procession reaches the far wall, a great mirror reflects the banner's image back — an identical copy appears flush against the opposite wall. Between the real banner and its reflection, the hall is flooded with each row's house color, as though the family's standard has claimed every stone in its path.

Grid: The hall is exactly as wide and tall as before — no stones are added or removed. The darkness is banished entirely; every black stone is now draped in color. The original banner holds its place at the left wall, its twin gazes back from the right, and between them stretches an unbroken field of each row's own heraldic hue.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (5389 chars)

We need to infer rule. From description: left edge banner: each horizontal stripe distinct coat of arms, framed on both sides by that stripe's own house color. Input shows leftmost columns have pattern: each row has some colors (non-black) at left side, rest black. In output, the banner is duplicated on right side, and between them fill with each row's house color (the stripe's own house color). Also each stripe leaves a ribbon of its house color trailing behind it as it moves across.

Let's examine Example1.

Input grid size? Let's count columns: "G R G . . . . . . . . ." that's 12 columns. Rows 4.

Input rows:
Row1: G R G . . . . . . . . .
Row2: B R B . . . . . . . . .
Row3: G R G . . . . . . . . .
Row4: B G B . . . . . . . . .

So leftmost three columns have pattern: each row has three colored cells (non-black). The rest are black '.'.

Output rows (same size 12):
Row1: G R G G G G G G G G R G
Row2: B R B B B B B B B B R B
Row3: G R G G G G G G G G R G
Row4: B G B B B B B B B B G B

Observations: The leftmost three columns unchanged? Let's compare row1: first three are G R G same as input. Row2: B R B same. Row3: G R G same. Row4: B G B same. So original banner stays at left.

Rightmost three columns? Let's see row1 last three columns: positions 10,11,12? Actually output row1: columns: 1 G,2 R,3 G,4 G,5 G,6 G,7 G,8 G,9 G,10 G,11 R,12 G. So rightmost three are columns 10-12: G R G? Wait column 10 is G, 11 is R, 12 is G. That's same as leftmost three. So mirror copy appears at right side, but maybe reversed? Not reversed; it's same orientation.

Between left and right, there is a field of each row's house color. What is house color? For each row, maybe the color that frames the stripe? In description: each stripe framed on both sides by that stripe's own house color. In input, each row's stripe maybe has a central color? Let's examine row1: pattern G R G. The outer colors are G, middle is R. So house color maybe G? Because outer frames are G. Row2: B R B -> house color B. Row3 same as row1. Row4: B G B -> house color B (outer B). So between left and right, we see many G's for rows 1 and 3, many B's for rows 2 and 4. Indeed output row1 columns 4-9 are G (six G's). Row2 columns 4-9 are B (six B's). Row3 same as row1. Row4 same as row2.

Thus rule: For each row, take the leftmost non-black segment (the "banner stripe") which consists of some pattern where the first and last cells are same color (house color) and there may be other colors inside. Keep that unchanged at left. Then replicate the same segment at rightmost side (same orientation). Fill all cells between left segment and right segment with the house color (the outer color of the segment). The width of fill is such that total width unchanged.

Thus algorithm: Determine width of left segment (call w). Determine total columns C. For each row, copy left segment to columns C-w+1 .. C (rightmost). Fill columns w+1 .. C-w with house color (the first cell of left segment). That's it.

Now 

... [2395 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.80)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 386: "A stained-glass window unfolds its hidden reflections when t ... [30 chars truncated]" k=1 alts: 6/15 67e8384a

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input is a 3×3 grid of colored cells. Each cell contains one of several colors, with no special structural constraints — any arrangement of colors is valid.

Do: 1. Start with the original 3×3 grid. 2. Create a horizontal mirror (flip left-to-right) of the original and place it to the right, forming a 3×6 grid. 3. Create a vertical mirror (flip top-to-bottom) of that entire 3×6 grid and place it below, forming the final 6×6 output.

Grid: The input is always a 3×3 grid. The output is always a 6×6 grid — the input tiled with its reflections to produce four-fold symmetry (left-right and top-bottom).

Figurative Descriptions (16 variants)

"A stained-glass window unfolds its hidden reflections when the cathedral doors swing open."

120b: k=1

See: A small square of stained glass, three panes across and three down, each shard holding its own color — a fragment of a larger window that has been folded in on itself, concealing its full design.

Do: The window remembers its complete form. First, the glass unfolds to the right, as though a hinged panel swings open to reveal the twin that was always pressed against it face-to-face. Then the entire doubled panel drops its reflection downward, like a window over still water, so that every shard finds its counterpart above and below. The cathedral reveals what symmetry had kept hidden.

Grid: The original fragment is a modest 3×3 piece of stained glass. When the cathedral fully opens its doors and the reflection pool settles beneath, the complete rose window stands 6×6 — four copies of the fragment facing inward like hands pressed together in prayer, symmetric in every direction.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (3018 chars)

We need to understand transformation. Input 3x3, output 6x6. The description: unfold to the right (mirror horizontally?) then reflection downward (mirror vertically). Essentially create a 2x2 tiling of the original and its mirrored versions? Let's examine example.

Input:
Row1: R X R
Row2: R M Y
Row3: R R R

Output 6x6:
Row1: R X R R X R
Row2: R M Y Y M R
Row3: R R R R R R
Row4: R R R R R R
Row5: R M Y Y M R
Row6: R X R R X R

Let's see pattern. The top-left 3x3 of output equals input? Compare:
Output rows1-3 cols1-3:
R X R
R M Y
R R R -> matches input exactly.

Now top-right 3x3 (rows1-3 cols4-6):
R X R
Y M R
R R R

Is that horizontal mirror of input? Input mirrored horizontally (left-right) would be:
R X R (same because symmetric?) Actually input row1: R X R reversed is R X R same. Row2: R M Y reversed is Y M R. Row3: R R R reversed same. So top-right matches that: row1 same, row2 Y M R, row3 same. Yes.

Now bottom-left 3x3 (rows4-6 cols1-3):
R R R
R M Y
R X R

That's vertical mirror (up-down) of input? Input vertical mirror (top-bottom) would be rows reversed: row3, row2, row1:
R R R
R M Y
R X R matches bottom-left.

Bottom-right 3x3 (rows4-6 cols4-6):
R R R
Y M R
R X R

That's both horizontal and vertical mirror (i.e., 180 rotation) of input. Indeed.

Thus transformation: create a 2x2 block where each quadrant is a copy of the input, mirrored appropriately: top-left original, top-right horizontal mirror, bottom-left vertical mirror, bottom-right both mirrors (i.e., rotate 180). Equivalent to reflecting across both axes to fill 6x6.

Thus rule: output size double each dimension, fill with input and its reflections.

Now apply to test input:

Test Input:
B M M
X R R
R R R

Let's denote matrix A (3x3). We'll compute:

A:
row1: B M M
row2: X R R
row3: R R R

Horizontal mirror (flip left-right) H:
row1 reversed: M M B
row2: R R X
row3: R R R (same)

Vertical mirror (flip up-down) V:
rows reversed:
row1: R R R
row2: X R R
row3: B M M

Both (HV) i.e., rotate 180 (or flip both): reverse rows and reverse each row:
Take A, reverse rows order and each row reversed:
Original rows reversed: row3, row2, row1 => R R R, X R R, B M M then each reversed:
row3 reversed: R R R (same)
row2 reversed: R R X
row1 reversed: M M B

Thus HV matrix:
row1: R R R
row2: R R X
row3: M M B

Now construct output 6x6:

Top-left (TL) = A
Top-right (TR) = H
Bottom-left (BL) = V
Bottom-right (BR) = HV

Let's write rows:

Rows 1-3 combine TL and TR side by side.

Row1 TL: B M M ; TR: M M B => combined: B M M M M B
Row2 TL: X R R ; TR: R R X => X R R R R X
Row3 TL: R R R ; TR: R R R => R R R R R R

Rows 4-6 combine BL and BR.

Row4 BL: R R R ; BR: R R R => R R R R R R
Row5 BL: X R R ; BR: R R X => X R R R R X
Row6 BL: B M M ; BR: M M B => B M M M M B

Thus output grid 6x6.

Check symmetry: matches pattern.

Now produce answer block with spaces.

Answer:

B M M M M B
X R R R R X
R R R R R R
R R R R R R
X R R R R X
B M M M M B

Let's ensure using color key let

... [42 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.25)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 387: "Two puzzle pieces, cast from the same mold, find each other ... [17 chars truncated]" k=1 alts: 10/15 681b3aeb

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains exactly two non-black colored shapes on a black background. Each shape is a distinct color. One shape is larger (occupies more cells) and the other is smaller. Both shapes fit within a 3×3 bounding box. The larger shape has empty (black) cells within its 3×3 bounding box that form a cavity, and the smaller shape's colored cells exactly correspond to that cavity pattern.

Do: 1. Identify the two colored shapes in the input grid. 2. Determine each shape's bounding box. Both will fit within a 3×3 region. 3. Extract each shape into a 3×3 grid, recording which cells are filled (colored) and which are empty (black). 4. Identify the larger shape (the one with more colored cells). 5. Overlay the two shapes into a single 3×3 output grid: place the larger shape's color in all cells it occupies, and place the smaller shape's color in the remaining cells (which are the empty/cavity cells of the larger shape). The smaller shape's filled cells exactly fill the gaps of the larger shape.

Grid: The output grid is always 3×3. Each cell contains one of the two colors from the input shapes. There are no black cells in the output — the two shapes together completely tile the 3×3 grid.

Figurative Descriptions (16 variants)

"Two puzzle pieces, cast from the same mold, find each other and become whole."

120b: k=1

See: Two lonely fragments drift in a vast dark ocean, each a different hue. One is the bigger sibling — nearly complete but riddled with holes, like a key missing its teeth. The other is the smaller sibling — just a few bright cells, shaped exactly like what the bigger one is missing. They are two halves of the same breath.

Do: The smaller fragment hears the larger one calling and swims into its wounds. Every gap in the bigger sibling is a socket waiting for the smaller one's bones. Nest them together: the larger fragment keeps its color wherever it was whole, and the smaller fragment fills every void with its own color, like mortar between bricks. They fuse into one solid body, no darkness left between them.

Grid: The fused body is a compact 3×3 tile — fully healed, no black scars remaining. Every cell glows with one of the two original colors: the dominant sibling's color where it was strong, the smaller sibling's color where it was needed.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10070 chars)

We need to infer rule from example. Input is 10x10 grid with many black '.' and some colored cells. Output is 3x3 grid. The description: larger fragment (bigger sibling) is nearly complete but riddled with holes, like a key missing its teeth. Smaller fragment is just a few bright cells, shaped exactly like what the larger one is missing. They are two halves of the same breath. The smaller fragment fills the gaps.

Thus we need to find two shapes: one larger shape of one color, mostly solid but with holes (black cells) that correspond to shape of smaller fragment of another color. Then output is the combined shape (3x3) where larger color fills its original cells, and smaller color fills the holes.

In example 1: Input shows M (magenta) and Y (yellow). Let's map coordinates. 10x10 grid. Let's locate M cells: rows (0-index). Row0: none. Row1: none. Row2: columns 3,4,5 are M. Row3: columns 3,4 are M. Row4: column5 is M. That's a shape like a small L? Let's draw 5x? Actually shape: positions:
(2,3)(2,4)(2,5)
(3,3)(3,4)
(4,5)
That's like a shape with missing cells at (3,5) and (4,3)(4,4). The Y cells: row0 col8? Wait Y at (0,8). Row1 col8,9 Y Y. So Y shape is a 2x2 block? Actually coordinates: (0,8) Y, (1,8) Y, (1,9) Y. That's like an L shape of three cells: a 2x2 missing one corner (top-left?). Let's see shape: cells at (0,8), (1,8), (1,9). That's like a small L shape (like missing (0,9)). The description says smaller fragment shaped exactly like what the larger one is missing. So the missing cells in M shape should match Y shape (maybe after translation/rotation?). Let's compute missing cells in M shape within its bounding box? M shape bounding box rows 2-4, cols 3-5 (size 3x3). Within that 3x3, cells present: (2,3)(2,4)(2,5); (3,3)(3,4); (4,5). Missing cells: (3,5), (4,3), (4,4). That's three cells forming an L shape (like Y shape). Indeed Y shape is three cells L shape. So Y shape matches missing cells shape.

Thus output 3x3 tile corresponds to the bounding box of larger shape (3x3). Fill with M where M existed, Y where missing. Output given:
M M M
Y M M
Y Y M

Let's map: rows correspond to bounding box rows 2-4, cols 3-5. Fill: row2: M M M matches. row3: Y M M (since missing at (3,5) becomes Y). row4: Y Y M (missing at (4,3),(4,4) become Y). Yes matches.

Thus rule: Identify the larger shape (color that appears more cells within its minimal bounding rectangle) that is mostly filled except some missing cells. Identify the smaller shape (different color) that matches the pattern of missing cells (same shape). Then output the combined 3x3 (or whatever size) rectangle of the larger shape's bounding box, with missing cells filled by smaller color.

Now test input: we have colors R (red) and A (azure). Let's locate them.

Input grid 10x10.

Rows index 0-9.

Row0 all . 
Row1 all .
Row2: ". . . . R . . . . ." => R at (2,4)
Row3: ". . . R R . . . . ." => R at (3,3) and (3,4)
Row4: all .
Row5: all .
Row6: ". . . . . . . A A ." => A at (6,7),(6,8)

... [7099 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 398: "Three painters compete for a single canvas, and the last to ... [33 chars truncated]" k=1 alts: 8/15 6a11f6da

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: The input is a tall grid consisting of three equally-sized square layers stacked vertically. Each layer is 5 rows by 5 columns. The top layer contains cells that are either blue (B) or black (.). The middle layer contains cells that are either azure (A) or black (.). The bottom layer contains cells that are either magenta (M) or black (.). Each layer acts as a binary mask for its respective color, where the colored cells are 'present' and black cells are 'absent.'

Do: Divide the input into three equal layers from top to bottom: the B-layer (blue), the A-layer (azure), and the M-layer (magenta). Each layer has the same dimensions. Create an output grid of the same dimensions as one layer. For each cell position, apply a priority rule to determine the output color: if the M-layer has a colored cell at that position, the output is M (magenta); otherwise, if the B-layer has a colored cell at that position, the output is B (blue); otherwise, if the A-layer has a colored cell at that position, the output is A (azure); otherwise, the output is black (.). In short, the priority order is M > B > A > black.

Grid: The input grid is 15 rows by 5 columns, composed of three stacked 5×5 layers. The output grid is 5 rows by 5 columns. The background color is black (.). The three colors used are blue (B), azure (A), and magenta (M), each exclusive to its respective layer in the input but potentially co-occurring at the same position in the output resolution.

Figurative Descriptions (16 variants)

"Three painters compete for a single canvas, and the last to arrive paints over everyone else."

120b: k=1

See: Three artists share a studio, each assigned a floor in a tall building. The blue painter works on the top floor, the azure painter on the middle floor, and the magenta painter in the basement. Each has sketched out where they want to place brushstrokes on an identical canvas — some cells filled, some left bare.

Do: The three painters must merge their plans onto a single shared canvas. The magenta painter, working deepest underground, has the strongest pigment — wherever she has marked a stroke, her color dominates and nobody can paint over it. The blue painter has the next strongest hand; his marks show through wherever magenta hasn't already claimed the spot. The azure painter is the gentlest — her color only appears in the gaps the other two left untouched. Where no artist wanted to paint at all, the canvas stays bare and dark.

Grid: The tall building has three identical floors stacked vertically, each the size of the final canvas. Each floor's sketch uses only one color against darkness. The shared canvas is a single floor's worth of space, where the strongest pigments rise from below to claim their territory.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (7259 chars)

We need to infer rule. The description: three floors (blue top, azure middle, magenta bottom). Input grid seems to be 15 rows? Actually example input is 15 rows, each row length 5? Let's examine example 1 input: lines:

". B B B B"
". B B . ."
". B B B ."
". . B B B"
". B B B ."
". A A . ."
"A . . A ."
". A . . A"
". . A . ."
"A . A . A"
". M . M M"
". . M M M"
". M . . ."
". M M . M"
". . . . ."

That's 15 rows, 5 columns. So three floors each of 5x5? Actually 15 rows = 3*5 rows. So each floor is 5 rows stacked vertically: top floor rows 0-4 (blue), middle rows 5-9 (azure), bottom rows 10-14 (magenta). Output is 5x5 grid (single floor). The rule: combine layers with priority: magenta > blue > azure? Wait description: magenta strongest (bottom), blue next, azure gentle (top). So output cell color determined by: if magenta layer has a colored cell (non-.) then output that color (M). Else if blue layer has colored cell then output that color (B). Else if azure layer has colored cell then output that color (A). Else '.' (black). However example output includes also A and B and M. Let's test with example 1.

Let's compute manually for a few cells.

We need to map coordinates: rows 0-4 blue, 5-9 azure, 10-14 magenta.

Take column 0 (first column). Input rows:

0: '.' (blue)
1: '.' (blue)
2: '.' (blue)
3: '.' (blue)
4: '.' (blue)
5: '.' (azure)
6: 'A' (azure)
7: '.' (azure)
8: '.' (azure)
9: 'A' (azure)
10: '.' (magenta)
11: '.' (magenta)
12: '.' (magenta)
13: '.' (magenta)
14: '.' (magenta)

Now output row0 col0? Output grid first row: ". M B M M". Wait that's row0 of output. Column0 char '.' . According to rule, magenta at row0? Need to map which row of output corresponds to which column? Actually output is 5x5, each cell corresponds to column across floors? The stacking: each floor is same size; we need to combine per cell across floors (same column and row index within floor). So output row i corresponds to row i within each floor (i from 0 to 4). So for output row0 col0, we look at blue row0 col0, azure row5 col0, magenta row10 col0. Those are '.' '.' '.' => output '.' matches.

Now output row0 col1 is 'M'. Let's check column1 row0: blue row0 col1 = 'B' (since row0 ". B B B B": col1 B). azure row5 col1 = '.' (row5 ". A A . .": col1 A? Wait row5 is ". A A . .": columns: 0 '.',1 'A',2 'A',3 '.',4 '.' . So col1 = 'A'. magenta row10 col1 = 'M' (row10 ". M . M M": col1 = 'M'). Since magenta present, output 'M' matches.

Row0 col2 output 'B'. Check: blue row0 col2 = 'B', azure row5 col2 = 'A', magenta row10 col2 = '.' . Magenta none, blue present, output should be blue 'B' (since blue stronger than azure). Indeed output 'B'.

Row0 col3 output 'M'. Blue row0 col3 = 'B', azure row5 col3 = '.', magenta row10 col3 = 'M' => M.

Row0 col4 output 'M'. Blue row0 col4 = 'B', azure row5 col4='.', magenta row10 col4='M' => M.

Matches rule.

Check another cell: output row1 col0 = 'A' (from output). Row1 corresponds to row index 1 within floor. Blue 

... [4457 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.04)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 400: "A cocoon dissolves, and its pigment is summoned inward only ... [27 chars truncated]" k=1 alts: 8/15 6a980be1

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has three structural layers: (1) A single-cell-thick border frame of a uniform color (the 'frame color' — azure, yellow, magenta, or blue depending on the puzzle instance) forming the outermost row/column on all four sides. (2) A green (G) spine consisting of two parallel green lines running either vertically or horizontally through the interior, dividing the grid into an interior band (between the green lines) and exterior regions (outside the green lines but inside the border). (3) Inside the spine band, a repeating pattern of red (R) and black (.) cells. The exterior regions (between the spine and the border) are all black.

Do: 1. Identify the frame color (the uniform color forming the 1-cell border on all four sides). 2. Identify the green spine — two parallel green lines running in the same direction (both horizontal or both vertical). 3. Note the pattern of red and black cells inside the spine band. 4. Build the output grid (same dimensions as input): (a) Every cell on a green spine line becomes G, extended across the full width or height of the grid. (b) For each row (if spine is vertical) or column (if spine is horizontal), examine what color appears in the spine interior for that row/column. If the spine interior contains R for that row/column, fill all non-green cells in that row/column with R (inside the spine) and the frame color (outside the spine). If the spine interior contains black (.) for that row/column, fill all non-green cells with black (.). 5. The border frame is removed — it does not appear in the output. The frame color only appears in the exterior regions of rows/columns whose spine interior contains red.

Grid: The output grid has the same dimensions as the input grid. There is no border frame. The green spine lines span the full extent of the grid. The grid is filled according to the spine's red/black pattern: red rows or columns get the former frame color outside the spine, and black rows or columns remain entirely black outside the spine.

Figurative Descriptions (16 variants)

"A cocoon dissolves, and its pigment is summoned inward only where the heartbeat pulses."

120b: k=1

See: A chrysalis of uniform color wraps the entire grid — a single-cell shell holding everything in. Inside, two parallel green veins run like arteries through the body, forming a narrow channel. Within that channel, a heartbeat pattern alternates: some segments pulse red, others lie dormant and dark. The flesh between the veins and the shell is entirely dark — asleep, waiting.

Do: The chrysalis breaks open and vanishes. But its color doesn't disappear — it becomes a dye that the red pulses can summon. Each red heartbeat in the channel calls out across its entire row or column, flooding the flesh beyond the veins with the old shell color, as if the pulse radiates outward through the green walls. Where the channel is dark and still, no signal is sent — those rows or columns remain black and silent from wall to wall. The green veins themselves stretch to span the full body, no longer caged by the shell. Think of it as: red pulses broadcast the cocoon's color outward; silence broadcasts nothing.

Grid: The output is the same size as the input. The shell is gone. The green veins now run edge to edge. Rows or columns with a red pulse in the channel are vivid — the old shell color fills everything outside the veins, red fills between them. Rows or columns where the channel was dark remain entirely black outside the veins. The grid breathes in a rhythm dictated by the heartbeat pattern.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (14658 chars)

We need to understand rule from example.

Input grid 19x19? Let's count: Example 1 input appears 19x19 with border of A (azure) around. Inside there is a vertical green column? Actually pattern: column 7? Let's map coordinates: each row has 19 entries. The border rows (first and last) are all A. The leftmost and rightmost columns are A. Inside, there is a vertical line of G at column 7? Let's see row2: "A . . . . . G R R R G . . . . . . . A". So after A, six '.' then G, then R R R, then G, then rest '.' then A. So column indices: 1 A, 2 .,3 .,4 .,5 .,6 .,7 .,8 G? Wait count: after A (col1), there are six '.' (col2-7), then G at col8, then R at col9,10,11, then G at col12, then '.' rest col13-? Let's count total 19. So green column at col8 and col12? Actually two green columns forming veins? In example, there are two parallel green veins vertical? Yes they are at columns 8 and 12 (distance 4 apart). The channel between them includes three columns (9-11) where there may be R or '.'.

The pattern: In some rows, the channel (cols9-11) is filled with RRR (red heartbeat) else '.' (dark). In example, rows 2-4 have RRR, rows5-7 have '.' (dark), rows8-10 have RRR, rows11-13 '.' etc. So alternating groups of three rows? Actually pattern: groups of 3 rows with RRR, then 3 rows with '.' etc. So heartbeat pattern vertical? Actually rows.

Now output: The border A disappears (becomes '.'?) In output first row is all '.' (black). Indeed shell color (A) becomes dye used to fill outside of veins where red present. The output shows that rows that had RRR in channel become filled with A (the old shell color) outside the green veins, while rows with '.' remain '.' outside.

Specifically output row2: "A A A A A A G R R R G A A A A A A A A". So outside the two green columns (col8 and col12) the cells become A (the shell color) for that row because there was red in channel. The green columns remain G, red remain R. The rest of grid (outside) becomes A for those rows. For rows without red, output row5: ". . . . . . G . . . G . . . . . . . ." i.e., outside remains '.' (black). So rule: For each row, if any red present in the channel (between the two green veins), then fill entire row (outside the veins) with the shell color (the color that formed the border). If no red, keep black.

Similarly columns? The description mentions rows or columns with red pulse broadcast across its entire row or column. In example, the veins are vertical, so channel is vertical? Actually red pulses are in rows, but they broadcast across row. Could also be columns if veins horizontal. So rule: Determine the two parallel green lines (veins) that are continuous across entire grid after transformation. They may be vertical or horizontal. The region between them is the channel. For each line (row or column) that contains at least one red cell within the channel, fill the entire line outside the veins with the shell color (the color that formed the outer border). For lines without red, keep black.


... [11667 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 408: "A lighthouse sends its beams to the horizon in four directio ... [3 chars truncated]" k=2 alts: 3/15 6cdd2623

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains scattered cells of several colors on a black background. Exactly one color — the 'marker' color — appears precisely four times, positioned at the endpoints of a cross pattern: two cells sit on opposite edges of the same row (leftmost and rightmost columns, or near them), and two cells sit on opposite edges of the same column (topmost and bottommost rows, or near them). All other non-black cells are noise/distractors.

Do: 1. Identify the unique color that appears exactly four times in the grid. This is the marker color. 2. Among those four marker positions, find the two that share the same row — they define the 'marker row.' Find the two that share the same column — they define the 'marker column.' 3. Fill the entire marker row with the marker color (every cell in that row becomes the marker color). 4. Fill the entire marker column with the marker color (every cell in that column becomes the marker color). 5. Set all other cells in the grid to black.

Grid: The grid dimensions match the input exactly (they vary across instances). The background is black (.). The output contains exactly one fully filled row and one fully filled column of the marker color, forming a cross, with their intersection sharing the marker color. All other cells are black.

Figurative Descriptions (16 variants)

"A lighthouse sends its beams to the horizon in four directions."

120b: k=2

See: The grid is a dark ocean at night, littered with flotsam of many colors drifting aimlessly. Hidden among the debris are exactly four lighthouse keepers of the same uniform color, stationed at the far edges of the sea — two guarding opposite shores of a single shipping lane, and two watching over opposite ends of a single channel. All other colored flotsam is just wreckage to be cleared away.

Do: Find the four keepers wearing matching uniforms — they are the only color that appears exactly four times. Two keepers face each other across the same horizontal shipping lane; two face each other along the same vertical channel. When the lighthouse activates, it sends a brilliant beam of that color flooding the entire shipping lane from coast to coast, and another beam blazing the full length of the channel from shore to shore. The light obliterates everything else — all flotsam sinks into the dark water.

Grid: The ocean keeps its original dimensions. Where the shipping lane and the channel cross, the beams merge into one blazing intersection. Every cell along those two lines glows with the keepers' color. The rest of the sea returns to total darkness.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (8974 chars)

We need to infer rule from examples.

We have grid with colors. The description: find color that appears exactly four times. Those four cells are at far edges: two opposite sides same row (horizontal lane) and two opposite sides same column (vertical channel). Then output grid: fill entire row (full width) and column (full height) with that color, intersect cell both. All other cells become '.' (black). In example1, which color appears exactly four times? Let's check input example1: colors present: A, B, R. Count occurrences: A appears many times? Let's count quickly: In input lines, there are many As. Not four. B appears many. R appears at positions: row4 col1? Actually row4 first cell R, and row4 last cell R. Row12 first cell R, last cell R. That's four Rs, located at far left/right edges of rows 4 and 12. They are at same rows (row4 and row12) and same columns (col1 and col20). So they form rectangle corners. Output: rows 4 and 12 become all R across entire row. Also columns 1 and 20 become all R across entire column. Indeed output shows row4 all R, row12 all R, column1 all R, column20 all R. Others '.'.

Thus rule matches.

Example2: Input colors: R, G, X, etc. Need color appearing exactly four times. Let's count: G appears many? In input, there are G at (row1 col6?), Actually row1: ". . . . . R R . . . . G . . . . . R . . . ." So G appears at column12 row1. Row9: "G . . X X . R X . . . . . . . . . . . . R G" includes G at col1 and col22? Wait width maybe 22? Let's count. Row9 has G at start and at end. Row11: ". . R . . . . . X . . G . . . . . X . X . ." includes G at column12. So total G count: row1 col12, row9 col1, row9 col22, row11 col12 = 4. They are at far edges? Row9 col1 is left edge, row9 col22 is right edge (same row). Row1 col12 is top edge? Not top edge (row1 is top row). Column12 is not edge column (since width maybe 22). Row11 col12 is bottom? Row11 is not bottom row (there are 11 rows). So not edges. But description says keepers at far edges of sea — two guarding opposite shores of a single shipping lane (horizontal), and two watching opposite ends of a single channel (vertical). In example2, the four G's are at positions: (row1, col12) top edge, (row9, col1) left edge, (row9, col22) right edge, (row11, col12) bottom edge. So they form a cross: same column (col12) top and bottom edges, same row (row9) left and right edges. Indeed they are at far edges of grid: top row, bottom row, leftmost column, rightmost column. So they define a horizontal lane at row9 (full row) and vertical channel at column12 (full column). Output shows column12 all G, row9 all G. Indeed output grid: column12 (index?) all G, row9 all G. Yes.

Thus rule: find color that appears exactly four times. Those four cells will be at the four extreme positions: two share same row (leftmost and rightmost columns), two share same column (topmost and bottommost rows). Then output: fill entire row (the row of the two cells) and column (the column of the other tw

... [6112 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 416: "A conductor reads the harmony of an orchestra to choose the ... [23 chars truncated]" k=2 alts: 15/15 6e02f1e3

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: The input is a 3x3 grid filled with colored cells. The cells may all be the same color, or they may contain two or three distinct colors arranged in any pattern.

Do: Count the number of distinct colors present in the input grid. Create a 3x3 output grid filled with black (.). Then place exactly three grey (X) cells according to the count: If there is 1 distinct color, place X across the entire top row (positions (0,0), (0,1), (0,2)) forming a horizontal line. If there are 2 distinct colors, place X along the main diagonal (positions (0,0), (1,1), (2,2)) running from top-left to bottom-right. If there are 3 distinct colors, place X along the anti-diagonal (positions (0,2), (1,1), (2,0)) running from top-right to bottom-left.

Grid: Both input and output are 3x3 grids. The output grid has a black (.) background with exactly three grey (X) cells arranged in one of three fixed patterns depending on the color count in the input.

Figurative Descriptions (16 variants)

"A conductor reads the harmony of an orchestra to choose the curtain call formation."

120b: k=2

See: A small stage is filled with performers dressed in colored costumes. Sometimes the entire ensemble wears matching outfits; other times, factions in different colors share the stage.

Do: After the performance, the curtain falls and the stage goes dark. A spotlight operator counts how many distinct factions performed. The fewer the factions, the more unified the bow — a single troupe lines up shoulder to shoulder across the front of the stage. A duo takes a bow along the grand staircase from balcony to pit. A trio crosses paths, saluting from opposite wings through center stage.

Grid: The dark stage is the same size as the original. Exactly three spotlights (grey) illuminate the bowing positions against the blackout, their arrangement a silent review of the evening's diversity.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (2201 chars)

We need to infer rule from examples.

We have grid of colors. Output contains X (grey) and . (black) only? In examples output only X and . . Yes.

Goal: count distinct colors (factions) in input? Then place exactly three X's (spotlights) in output arrangement based on number of distinct colors? Let's examine.

Example1 input: all G (green) only one color. Output:
X X X
. . .
. . .
So three X's in first row, rest . . . So arrangement maybe a line across front (top row) when only one faction? "single troupe lines up shoulder to shoulder across the front of the stage." That matches: a single line of X across front (top row). So three X's across top row.

Example2 input: colors: G, Y, R. Let's list distinct: G, Y, R = 3 colors. Output:
. . X
. X .
X . .
That's a diagonal from bottom-left to top-right? Actually positions: (row1,col3), (row2,col2), (row3,col1). That's a diagonal descending leftwards (anti-diagonal). That's three X's forming a line from top right to bottom left. That could represent "a duo takes a bow along the grand staircase from balcony to pit" maybe two? Not sure.

But we have three X's always exactly three. Arrangement varies based on number of distinct colors? Example1 (1 distinct) gave top row. Example2 (3 distinct) gave anti-diagonal. Need mapping for 2 distinct maybe? Not given. For test input we have colors: Y, R, G. That's also three distinct colors (Y,R,G). Same as example2. So output likely same pattern: anti-diagonal.

But need verify if arrangement depends on something else like which colors present? Example2 input grid:
G G G
Y R R
Y Y R

Let's compute something: maybe count of each color? Not needed.

Maybe rule: place X's at positions of the three most frequent colors? Not.

But output pattern independent of input arrangement? Example1 all same color gave top row. Example2 mixed gave anti-diagonal. Could be based on number of distinct colors: 1 -> top row, 2 -> middle column? 3 -> anti-diagonal? Need to guess for 3 distinct we output anti-diagonal.

Thus test input also 3 distinct, output same anti-diagonal.

Let's produce grid same size as input (3x3). So output:

Row1: . . X
Row2: . X .
Row3: X . .

Same as example2.

Let's answer.

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 417: "A boomerang's throw traces the arc of its missing wing" k=1 alts: 10/15 6e19193c

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input grid contains one or more L-shaped trominos (groups of exactly 3 colored cells) on a black background. Each tromino occupies 3 out of 4 cells of a 2×2 bounding box, forming an L-shape. The 4th corner of that 2×2 box is empty (black).

Do: For each L-shaped tromino, identify its 2×2 bounding box and locate the missing (empty) 4th corner. Determine the diagonal direction that points away from the center of the L-shape (i.e., outward from the missing corner, away from the occupied cells). Starting from that missing corner, draw a diagonal line of colored cells (using the same color as the tromino) extending one cell at a time in that direction until reaching the edge of the grid. Do not modify the original tromino cells.

Grid: The grid dimensions match the input (typically 10×10). The background is black (.). The coordinate system is row-column with (0,0) at the top-left. All original cells are preserved; only new diagonal trail cells are added.

Figurative Descriptions (16 variants)

"A boomerang's throw traces the arc of its missing wing"

120b: k=1

See: On a dark field, small boomerangs rest in different orientations — each one a colored L-shape missing a single wing tip, as if snapped off before flight.

Do: Each broken boomerang launches from its missing wing tip. The ghost of the lost piece flies outward along the diagonal it would have traced, leaving a streak of color behind it as it sails away until it vanishes beyond the edge of the field. The boomerang itself stays planted where it lies.

Grid: The dark field keeps its original size. The boomerangs remain exactly where they fell. Only the flight trails — colored streaks from each missing tip — are new marks on the field.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (5821 chars)

We need to infer rule. Description: boomerangs are L-shape missing a tip. Each missing tip launches a colored streak along diagonal outward.

We need to identify L-shape pattern: In example 1 input, there are W's forming small L shapes missing a tip. Let's map coordinates. Grid 10x10.

Let's index rows 0-9, cols 0-9.

Input example1 positions of W:
(1,3),(1,4)
(2,4)
(6,3)
(7,3),(7,4)

These look like two L shapes? First cluster: (1,3),(1,4),(2,4) forms an L shape missing (2,3) tip? That's a 2x2 block missing bottom-left. That's like a boomerang missing tip at (2,3). The missing tip is diagonal from the corner? The L shape orientation: The missing tip is the cell that would complete the 2x2 square. So missing tip is at (2,3). According to description, ghost flies outward along diagonal from missing tip. Which diagonal? The diagonal that would have been traced by the missing piece? For a missing bottom-left cell, the diagonal direction outward would be down-left? Actually from missing tip (2,3) outward away from the L shape? Let's see output: Additional W's appear along a diagonal line from that missing tip outward until edge.

Output has extra W's at positions:
(1,9?) Wait output row1 (index1) has W at col9? Let's list output W positions:

Row1 (index1): columns 3,4,9? Actually output row1: ". . . W W . . . . W" => cols3,4,9.
Row2: ". . . . W . . . W ." => cols4,8.
Row3: ". . W . . . . W . ." => cols2,7.
Row4: ". W . . . . W . . ." => cols1,6.
Row5: "W . . . . W . . . ." => cols0,5.
Row6: ". . . W . . . . . ." => col3.
Row7: ". . . W W . . . . ." => cols3,4.
Rows8-9 unchanged.

So new W's added form a diagonal line from missing tip (2,3) moving down-left? Let's see coordinates: missing tip (2,3). Down-left diagonal would be (3,2),(4,1),(5,0) then out. Indeed we have W at (3,2) row3 col2? Output row3 (index3) has W at col2 yes. Row4 col1 yes. Row5 col0 yes. That's the diagonal down-left. Also there is a diagonal up-right? Not present. So only one direction.

What about second L shape at (6,3),(7,3),(7,4) missing (6,4) (top-right missing). Missing tip at (6,4). Diagonal outward from missing tip would be up-right? Actually missing top-right, outward direction would be up-right (north-east). Let's see added W's: from (6,4) go up-right: (5,5),(4,6),(3,7),(2,8),(1,9). Indeed output has W at (5,5) row5 col5 yes, (4,6) row4 col6 yes, (3,7) row3 col7 yes, (2,8) row2 col8 yes, (1,9) row1 col9 yes. So matches.

Thus rule: Identify each L-shaped pattern of three same-colored cells forming a 2x2 block missing one cell (the "missing tip"). For each such shape, extend a line of that color from the missing cell outward along the diagonal away from the existing cells (i.e., direction from the missing cell opposite the interior of the L). Fill cells along that diagonal until edge, leaving original cells unchanged.

Now test input with A's.

Let's locate A positions.

Rows:

Row0 all .
Row1 all .
Row2: col4 A (since ". . . . A . . . . .") index2

... [2857 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 418: "A naturalist classifies unknown species by the size of their ... [7 chars truncated]" k=1 alts: 6/15 6e82a1ae

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains several distinct clusters (connected components) of grey (X) cells against a black (.) background. Each cluster varies in size, meaning the number of cells it contains. Clusters can be any shape — L-shapes, straight lines, squares, etc. — and are scattered across the grid without overlapping.

Do: 1. Identify every connected component of grey (X) cells in the input grid, using 4-connectivity (up, down, left, right — not diagonal). 2. Count the number of cells in each connected component. 3. Recolor each component based on its size: cells belonging to a 2-cell cluster become green (G), cells belonging to a 3-cell cluster become red (R), and cells belonging to a 4-cell cluster become blue (B). 4. All black (.) background cells remain unchanged.

Grid: The grid dimensions match the input exactly (typically 10x10). The background is black (.). The only non-background elements are clusters of grey cells in the input, which become green, red, or blue in the output depending on cluster size. No cells are added or removed; only colors change.

Figurative Descriptions (16 variants)

"A naturalist classifies unknown species by the size of their herds."

120b: k=1

See: Scattered across a dark landscape, small herds of mysterious grey animals huddle together. Some travel in pairs, some in trios, and some in groups of four. Each herd keeps to itself, never touching another.

Do: A naturalist arrives to catalog the wildlife. She identifies each herd by counting its members, then paints them with a tag color revealing their species: the smallest pairs share one color, the middle-sized trios another, and the largest bands of four a third. The empty wilderness stays untouched.

Grid: The dark terrain remains as-is. Every grey animal receives its species tag — a new coat color determined solely by how many companions share its herd. No creatures appear or vanish; only their coats change.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (7862 chars)

We need to infer rule: In example, input has X (grey) representing animals. They form clusters (connected components) of size 2,3,4? Let's examine example1 input grid 10x10.

I'll map coordinates. Let's list X positions:

Row2 (0-index? but easier). Let's index rows 0-9.

Row0: none
Row1: none
Row2: columns 1,2,3 are X X X (a horizontal line of 3). Also column7 is X.
Row3: column7 X.
Row4: none
Row5: columns3,4 X X (two adjacent horizontally)
Row6: column3 X
Row7: column7 X
Row8: columns1,2 X X ; column5 X ; columns7,8,9 X X X (three contiguous)
Row9: columns1,2 X X

Now clusters: Need to consider adjacency? Likely orthogonal (4-direction) connectivity. Let's identify clusters.

Cluster A: row2 col1-3 (three X contiguous horizontally). That's size 3.

Cluster B: row2 col7, row3 col7, row7 col7, row8 col7-9? Wait connectivity: row2 col7 adjacent to row3 col7 (vertical) => same cluster. Row3 col7 adjacent to row4? none. Row7 col7 is far (row difference 4) not connected because there is gap rows4-6 no X at col7. So separate cluster. Row8 col7-9 are three X contiguous horizontally, but they are at row8, col7 adjacent vertically to row7 col7 (since row7 col7 X). So they belong to same cluster as row2 col7? Actually there is a gap rows4-6 col7 empty, so not connected. So row7 col7 connects to row8 col7-9, forming cluster size 4? Let's count: row7 col7 (1), row8 col7,8,9 (3) => total 4. Also row2 col7 and row3 col7 are separate cluster size 2? Wait row2 col7 and row3 col7 are adjacent vertically, so cluster size 2. So we have cluster size 2 at rows2-3 col7.

Cluster C: row5 col3-4 (two adjacent) size2, plus row6 col3 (adjacent vertically to row5 col3) => that makes size3? Actually row5 col3 adjacent to row6 col3, row5 col4 adjacent to row5 col3. So cluster of three X: positions (5,3),(5,4),(6,3). That's size 3.

Cluster D: row8 col1-2 and row9 col1-2: that's a 2x2 block? Actually row8 col1,2 X; row9 col1,2 X. That's four X forming a 2x2 square, size 4.

Now output colors: Let's map.

Output grid shows at those positions colors: For cluster size 3 (row2 col1-3) output R R R (red). Indeed output row2 col1-3 are R. So size3 -> red.

Cluster size 2 (row2-3 col7) output? Row2 col7 in output is G? Wait output row2 col7 is G (green). Row3 col7 is G. So size2 -> green? But also other size2 clusters: row5 col3-4? Output row5 col3-4 are R R (red) actually row5 col3-4 are R R. That's size2 but colored red, not green. Hmm maybe orientation matters? Let's double-check mapping.

Hold on: In output row5 col3-4 are R R. In input those are part of cluster size3 (including row6 col3). So cluster size3 gave red. So that matches size3 -> red.

Now size2 cluster row2-3 col7 gave green. So size2 -> green? But there is also size2 cluster maybe elsewhere? Let's see other size2 clusters: none else.

Size4 clusters: row7-8 col7-9 (size4) output? Row7 col7 is B (blue). Row8 col7-9 are B B B. So size4 -> blue.

Size4 cluster row8-9 col1-2 (2x2) output? Row8

... [4925 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.46)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 422: "A kingdom builds its castle walls to protect the darkness wi ... [5 chars truncated]" k=3 alts: 7/15 6f8cd79b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input is a rectangular grid filled entirely with black (.) cells. The grid can be any width and height.

Do: Color every cell that lies on the border of the grid azure (A). A border cell is any cell in the first row, last row, first column, or last column. Leave all interior cells black (.).

Grid: The output grid has the same dimensions as the input grid. The background is black, with a 1-cell-thick azure frame around the entire perimeter.

Figurative Descriptions (16 variants)

"A kingdom builds its castle walls to protect the darkness within."

120b: k=3

See: A quiet kingdom of shadow stretches across the land — nothing but still, dark earth as far as the eye can see.

Do: The kingdom raises azure fortress walls along every exposed edge, one stone thick. The sentinels stand guard only where the land meets the outside world. The interior remains untouched — a sheltered courtyard of darkness the walls were built to protect.

Grid: The kingdom keeps its original shape. Azure stone rises wherever the realm touches the frontier, forming an unbroken perimeter. The dark heart within sleeps undisturbed.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (1368 chars)

We need to interpret puzzle. Input grids are all black '.' cells. Output adds azure 'A' border of thickness one around the shape, leaving interior unchanged (still black). The shape is the original shape (the region of '.'), but we add A on the perimeter cells that are adjacent to outside (i.e., cells that are on the outer boundary of the shape). In examples, input is rectangular block of '.' of size (height x width). Output is same size, but border cells become A, interior remain '.'.

Check Example1: Input 5 rows, 4 columns. Output: border A, interior '.' (3x2 interior). Yes.

Example2: Input 4x3, output border A.

Example3: Input 5x6, output border A.

Thus rule: For any input grid, replace all '.' cells that are on the outer edge of the shape (i.e., have at least one neighbor that is outside the shape or outside grid) with 'A', keep interior '.' unchanged. Since shape is whole grid (all '.'), the border is the outermost rows and columns.

Test input: 7 rows, 6 columns (looks like 7x6). So output should be 7x6 with border A, interior '.'.

Thus rows:

Row1: all A (6)
Row2: A then 4 '.' then A
Row3: same
Row4: same
Row5: same
Row6: same
Row7: all A

So produce grid lines with spaces.

Let's write:

A A A A A A
A . . . . A
A . . . . A
A . . . . A
A . . . . A
A . . . . A
A A A A A A

Check dimensions: 7 rows, 6 columns.

Return as answer block.

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.33)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 430: "Rockets on a launchpad ignite according to their fuel type." k=2 alts: 15/15 72a961c9

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains a single horizontal bar made of blue (B) cells spanning one full row, with one or more non-blue colored marker cells (such as azure A or red R) embedded at specific positions within that bar. The rest of the grid is black (.) empty space. The bar can appear at any row in the grid.

Do: For each non-blue marker cell in the horizontal bar, grow a vertical column upward from that marker's position. The column's height (including the marker copy at the top) depends on the marker's color: azure (A) produces a column of height 3 (the top cell is A, followed by 2 blue B cells below it, connecting down to the bar), and red (R) produces a column of height 4 (the top cell is R, followed by 3 blue B cells below it, connecting down to the bar). The marker cell in the original bar remains unchanged. All other cells remain black.

Grid: The grid dimensions match the input exactly and do not change. The background is black (.). The horizontal bar occupies exactly one row. Columns grow strictly upward from the bar; no cells are modified below the bar or beyond the grid boundaries.

Figurative Descriptions (16 variants)

"Rockets on a launchpad ignite according to their fuel type."

120b: k=2

See: A launchpad stretches across the darkness — a solid blue runway with special rocket engines embedded in it, each glowing a distinct color (azure or red) to signal what kind of fuel they carry.

Do: When launch is triggered, each colored engine fires a plume straight up into the night sky. The plume is a column of blue exhaust capped by a bright flare matching the engine's color. Red rockets burn hotter and their plumes reach higher than azure ones. The engines themselves stay bolted to the launchpad.

Grid: The night sky stays dark except where exhaust trails rise. The launchpad remains fixed in its row. Plumes only fire upward — nothing falls below the runway, and no rocket escapes the edge of the sky.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (7890 chars)

We need to infer rule. From description: launchpad is a solid blue runway (row of B) with special rocket engines embedded, each glowing distinct color (azure A or red R) to signal fuel. When launch triggered, each colored engine fires a plume straight up: column of blue exhaust capped by a bright flare matching engine's color. Red rockets burn hotter and plumes reach higher than azure ones. Engines stay in launchpad row. Plumes only fire upward, nothing below runway, no rocket escapes edge.

Interpretation: Input grid has a row of B (blue) representing runway. Some cells in that row are A (azure) or R (red) representing engines. Output adds vertical columns above each engine: a column of B cells (blue exhaust) extending upward some number of rows, ending with a cell of the engine's color (A or R) at the top of the column. For red rockets, plume reaches higher than azure ones (i.e., longer column). Need to determine lengths: maybe red plume length = number of rows above runway? In examples, we can deduce.

Example1: Input grid size? Let's count rows: Input shows 9 rows? Actually list: . . . . . . (row1) ... row2 row3 row4 B A B B B B (row4) then rows5-9 all dots. Output: rows1-9 with modifications. Let's map coordinates. Runway row is row4 (0-index maybe). In output, we see column above A (at column2?) Let's index columns 1-6. Input row4: B A B B B B. So A at column2. Output: row1: . A . . . . (so A at top row column2). Row2: . B . . . . (B at column2). Row3: . B . . . . (B at column2). Row4 unchanged runway. So plume: top cell A, then two B cells below it, then runway B at row4 column2? Actually runway at row4 column2 is A originally, not B. Wait runway row has A at column2, not B. So plume includes the engine cell itself? The description says engine stays in runway row. So output should keep A at runway row, not replaced. In output row4 column2 is A? Let's check output row4: B A B B B B (same). Yes row4 column2 is A. So plume above includes B cells directly above engine, then top cell is A (flare). In output, row1 column2 is A (top), rows2-3 column2 are B, row4 column2 is A (engine). So plume length for azure = 2 B cells above plus top A. So total height (including engine) = 3 rows? Actually engine row is part of runway, not counted as plume length. The plume extends upward: number of B cells = maybe count of B cells above engine equals number of B cells to left/right? Not.

Now Example2: Input grid: 10 rows? Let's count: rows: 1-5 all dots, row6 B B B R B B B (runway), rows7-10 all dots. Output: rows1-5 have modifications: row1 all dots, row2 . . . R . . . (R at column4 top), row3 . . . B . . . (B at column4), row4 . . . B . . ., row5 . . . B . . . . Row6 unchanged runway. So plume for red engine at column4: top cell R at row2, then three B cells rows3-5, then engine row6 column4 is R. So red plume has 3 B cells above, azure had 2 B cells above. So red plume longer by 1 row. So rule: azure plume height = 2 B cells above engine; red plume height

... [4890 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 433: "Raindrops fall toward a drumhead and splash upward on impact" k=1 alts: 6/15 73c3b0d8

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains a horizontal red line spanning the full width of the grid, a set of yellow dots scattered on the black background (some above the red line, some below it), and otherwise empty black cells.

Do: 1. Shift every yellow dot down by exactly one row. The red line remains fixed in place. 2. After shifting, check each yellow dot that is now positioned exactly one row above the red line. For each such dot, generate a V-shaped bounce pattern: place additional yellow dots along two diagonal paths extending upward from that dot's position — one going up-left (row-1, col-1 per step) and one going up-right (row-1, col+1 per step). Continue each diagonal path until it exits the grid boundary. 3. Yellow dots that are not exactly one row above the red line after shifting (whether above or below the red line) simply remain at their new shifted position with no bounce effect. 4. All other cells remain unchanged.

Grid: The grid dimensions and the position of the red horizontal line vary across instances. The background is black (.). The red line always spans the full width of the grid. Yellow dots may appear anywhere above or below the red line. The coordinate system is row-column with row 0 at the top.

Figurative Descriptions (16 variants)

"Raindrops fall toward a drumhead and splash upward on impact"

120b: k=1

See: Yellow raindrops hang suspended in a dark sky above and below a taut red drumhead stretched across the world. The air is still, everything frozen mid-fall.

Do: Gravity pulls each raindrop one step closer to the earth. The red drumhead is anchored and never moves. Any drop that lands right on the drumhead's edge erupts into a V-shaped splash — twin arcs of yellow spray flying up-left and up-right, each droplet one step further along the diagonal until it escapes the sky. Drops that are still drifting through open air, or that have already passed beneath the drum, simply settle into their new resting place — no splash, no fanfare.

Grid: The dark sky (black) can stretch to any height or width. The red drumhead always runs its full width like a horizon line, dividing above from below. Yellow raindrops may float anywhere in the sky or sit beneath the drum. The world is read top-to-bottom, left-to-right.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (12519 chars)

We need to infer rule. From description: red drumhead is a full-width row of R that never moves. Yellow raindrops (Y) fall one step down (gravity) unless they hit the drumhead edge? Actually "any drop that lands right on the drumhead's edge erupts into a V-shaped splash — twin arcs of yellow spray flying up-left and up-right, each droplet one step further along the diagonal until it escapes the sky." So if a Y is directly above a R cell? The drumhead edge maybe the boundary between R row and the row above? The "edge" maybe the cell directly above a R? When a Y moves down onto a R cell? But R cells are occupied, can't have Y there. So maybe when Y is directly above R, after moving down it would land on R, causing splash: the Y disappears and creates two Y's moving diagonally up-left and up-right from that point (the edge). They continue moving diagonally outward each step until they exit grid (i.e., go beyond top row). In example, we can test.

Let's parse example 1. Input grid size? Let's count rows: appears 12 rows? Actually list: rows: 1 . Y . . . . . . ; 2 . . . . . Y . . ; 3 . . . . . . . . ; 4 . . . . . . . . ;5 . . Y . . . . . ;6 . . . . . . . . ;7 . . . . . . . . ;8 . . . . Y . . . ;9 . . . . . . . . ;10 R R R R R R R R ;11 . . . Y . . . . ;12 . . . . . . . . . Wait row 12 maybe . . . . . . . . (but they show . . . . . . . .). Actually output also 12 rows.

So drumhead at row 10 (index 10). Y's above and below. In output, Y's have moved down one step (gravity) except those that hit drumhead? Let's see.

Input Y positions:
(1,2) row1 col2 (0-index?). Row2 col6. Row5 col3. Row8 col5. Row11 col4.

After transformation output Y positions:
Row2 col2? Actually output row2: ". Y . . . . . ." => Y at col2 row2 (was row1 col2 moved down 1). Good.
Row3: ". . . . . Y . ." => Y at col6 row3 (was row2 col6 moved down 1). Good.
Row5: "Y . . . . . . ." => Y at col1 row5 (was row5 col3? Wait row5 col3 moved down? Actually row5 col3 moved down to row6? Let's see output row6: ". Y Y . . . . Y". Row6 has Y at col2, col3, col8. Row5 has Y at col1. So original Y at row5 col3 moved down to row6 col3? But output row6 col3 is Y indeed. Also there is Y at row6 col2 maybe from something else (splash). And Y at row6 col8 maybe from splash.

Original Y at row8 col5 moved down to row9 col5? Output row9: ". . . . Y . . ." => Y at col5 row9. Yes.

Original Y at row11 col4 (below drumhead) moves down? Row11 is below drumhead (row10). Gravity still applies? It says drops that are still drifting through open air, or that have already passed beneath the drum, simply settle into their new resting place — no splash. So they also move down one step? Row11 col4 moves to row12 col4. Output row12: ". . . Y . . . ." => Y at col4 row12. Yes.

Now what about splash? Need to see extra Y's created: output row6 col2, row6 col8, row7 col3? Actually output row7: ". . Y . . . Y ." => Y at col3 and col7. Row8: ". . . Y . Y . ." => Y at col4 and col6. Row6 c

... [9573 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.73)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 441: "A stone dropped at a crossroads sends square ripples through ... [13 chars truncated]" k=1 alts: 12/15 759f3fd3

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input grid contains a green cross: one full horizontal line of green cells and one full vertical line of green cells that span the entire grid and intersect at a single point. All other cells are black. The green lines divide the grid into four rectangular quadrants.

Do: Keep all green cells unchanged. For every non-green cell at position (row, col), compute its distance to the horizontal green line as dr = |row - row_h| and its distance to the vertical green line as dc = |col - col_v|, where row_h is the row index of the horizontal green line and col_v is the column index of the vertical green line. Then compute max(dr, dc). If this value is even (including zero), color the cell yellow. If this value is odd, leave the cell black. This produces nested concentric rectangular rings (based on Chebyshev distance) radiating outward from the green cross intersection.

Grid: The output grid has the same dimensions as the input. The green cross remains in place. Every non-green cell is either yellow or black depending on the parity of its Chebyshev distance (max of horizontal and vertical distances) from the cross. The pattern is symmetric across both the horizontal and vertical green lines.

Figurative Descriptions (16 variants)

"A stone dropped at a crossroads sends square ripples through still water."

120b: k=1

See: Two green roads cross the world — one running horizon to horizon, the other pole to pole — meeting at a single crossroads. The land around them lies dark and silent, divided into four quiet provinces.

Do: Where the roads meet, drop a stone into the darkness. Square-shaped ripples of golden light pulse outward from the crossroads, each ring one step larger than the last. But the light breathes — every other ripple is visible, and the ones between fade back to darkness. The roads themselves are ancient and unchanging; the ripples wash around them but never disturb them.

Grid: The green crossroads stands untouched. Golden ripple-rings nest outward like the walls of a square labyrinth, alternating with corridors of darkness between them. The pattern mirrors itself perfectly across both roads, as ripples naturally would from a single point of impact. The closer to the crossroads, the tighter the rings; the farther away, the wider each band of light and shadow.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (8784 chars)

We need to understand the puzzle. Input grid has green roads: one vertical line of G across entire grid, and one horizontal line of G across entire grid. In example 1, grid size 10x10. Input shows vertical line at column 6 (0-index?) Actually they show 10 columns. The vertical line of G in column 6 (index 5?) Let's index from 0. Input rows: each row has G at position 6 (1-indexed 7?). Wait row strings: ". . . . . . G . . ." That's 10 columns, G at column 7 (1-indexed). So vertical line at column 7. Horizontal line at row 5 (0-indexed 4?) Actually row 5 (0-indexed 4) is "G G G G G G G G G G". So horizontal line across row 5 (index 4). So they intersect at (row5, col7). That's the crossroads.

Output shows pattern of Y (yellow) forming square rings around the intersection, but not overlapping the green lines. Let's examine output grid 10x10.

Row0: "Y . Y Y Y Y G Y Y Y"
Row0 columns: col1 Y, col2 ., col3 Y, col4 Y, col5 Y, col6 Y, col7 G, col8 Y, col9 Y, col10 Y.

Row1: "Y . Y . . . G . . ."
Row2: "Y . Y . Y Y G Y Y ."
Row3: "Y . Y . Y . G . Y ."
Row4: "G G G G G G G G G G"
Row5: "Y . Y . Y . G . Y ."
Row6: "Y . Y . Y Y G Y Y ."
Row7: "Y . Y . . . G . . ."
Row8: "Y . Y Y Y Y G Y Y Y"
Row9: ". . . . . . G . . ."

Observations: The pattern of Y forms concentric squares centered at intersection (row4, col6? Actually intersection at row5? Wait row index 4 (0-based) is the horizontal line row. Intersection cell is row4 col6 (0-based). In output, that cell is G (unchanged). The Y pattern forms rings: distance Manhattan? Actually square rings (Chebyshev distance) from center, but only on cells not on the green lines. Also alternating rings: visible (Y) then invisible (.) then Y etc. Let's compute Chebyshev distance d = max(|r - rc|, |c - cc|). For each cell not on green lines, if d is even? Let's test: Intersection at (4,6). For cell (0,0) distance max(4,6)=6. 6 is even? 6 even => maybe Y? Output at (0,0) is Y. Good. Cell (0,1) distance max(4,5)=5 odd => output '.' (yes). Cell (0,2) distance max(4,4)=4 even => Y (yes). So pattern matches even distances Y, odd distances '.' (except green lines). However need to consider that green lines themselves remain G regardless of distance. Also note that cells on green lines but not intersection remain G (vertical line column6 all rows, horizontal line row4 all columns). In output they stay G.

Also note that the pattern is symmetric across both axes (roads). So rule: For each cell not G, set to Y if Chebyshev distance from intersection is even, else '.' (black). That matches example.

Now test input is larger: 12x12? Let's count columns: ". . . . . . G . . . . ." that's 12 columns. Rows: 12 rows. Horizontal line is row 9? Actually row index 8 (0-based) is "G G G G G G G G G G G G". So intersection at (row8, col7?) Wait vertical line column is column 7 (1-indexed 7?) Let's count: columns: positions 1-12. G appears at column 7 (since there are six '.' then G). So column index 6 (0-based). Intersection at (ro

... [6516 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 443: "A yellow weather vane tells the azure cloud which way to cas ... [17 chars truncated]" k=2 alts: 5/15 760b3cac

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a 6×9 grid with a black background containing two 3×3 regions, both occupying columns 3–5. The top region (rows 0–2) contains an azure pattern — an irregular shape made of azure cells within that 3×3 block. The bottom region (rows 3–5) contains a yellow pattern that resembles a plus sign (center cell plus four cardinal neighbors) with one additional cell in either the top-left or top-right corner of the 3×3 block. This extra yellow cell breaks the plus sign's symmetry and serves as a directional indicator: if it is in the top-right corner, the direction is RIGHT; if it is in the top-left corner, the direction is LEFT.

Do: 1. Identify the yellow 3×3 region (rows 3–5, columns 3–5). Determine whether the extra yellow cell (beyond the basic plus shape) is at the top-right or top-left corner of this block. Top-right means direction RIGHT; top-left means direction LEFT. 2. Extract the azure 3×3 pattern from rows 0–2, columns 3–5. 3. Create a horizontally mirrored copy of that 3×3 azure pattern (flip left to right, so column 0 of the block swaps with column 2). 4. Place the mirrored copy adjacent to the original azure block in the indicated direction: if RIGHT, place it at columns 6–8 (immediately to the right); if LEFT, place it at columns 0–2 (immediately to the left). 5. Leave the yellow region and all other cells unchanged.

Grid: The grid is always 6 rows by 9 columns with a black (.) background. The top three rows contain the azure pattern centered in columns 3–5, with columns 0–2 or 6–8 initially empty (black) to receive the mirrored copy. The bottom three rows contain the yellow directional indicator in columns 3–5. The grid dimensions do not change between input and output.

Figurative Descriptions (16 variants)

"A yellow weather vane tells the azure cloud which way to cast its reflection."

120b: k=2

See: A small azure cloud drifts in the upper half of a dark sky, while below it sits a yellow weather vane — a plus-shaped post with one extra fin jutting to the left or right, revealing which way the wind is blowing.

Do: The wind catches the azure cloud and blows a mirror-image of it outward in the direction the weather vane points. The reflection lands right next to the original cloud, as if the sky now holds the cloud and its twin reflected in a puddle laid sideways. The weather vane itself is anchored and never moves — only the cloud above is subject to the wind.

Grid: The sky is a dark canvas nine columns wide and six rows tall. The cloud always forms in the middle three columns of the upper atmosphere, while the weather vane is bolted into the same middle columns of the lower half. One side of the sky is empty, waiting to receive the cloud's wind-blown reflection.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (12917 chars)

We need to infer rule. Let's analyze examples.

We have 9x6 grid. Middle three columns are columns 4,5,6? Let's index 1-9. In inputs, azure A appears in columns 4 and 6 maybe? Let's view Example1 input rows:

Row1: . . . A . A . . . => A at col4 and col6.
Row2: . . . A A A . . . => A at col4,5,6.
Row3: . . . A A . . . . => A at col4,5.
Rows4-6 have Y pattern: plus shape with extra fin left or right? Row4: . . . . . Y . . . => Y at col6? Wait col6? Actually columns: col1-9. Row4 has Y at col6? Let's count: positions: 1 .,2 .,3 .,4 .,5 .,6 Y,7 .,8 .,9 . So Y at col6.
Row5: . . . Y Y Y . . . => Y at col4,5,6.
Row6: . . . . Y . . . . => Y at col5? Actually row6: . . . . Y . . . . => Y at col5.

So Y forms a plus shape centered at (row5,col5) with extra fin? Actually plus shape would be center plus up/down/left/right. Here we have Y at (row5,col5) center, up at row4,col5? but row4 has Y at col6 not col5. So it's shifted? Let's map coordinates: center maybe at (row5,col5). Up would be row4,col5 but there is no Y there. Instead Y at row4,col6 (right-up diagonal). Down row6,col5 is Y present. Left row5,col4 Y present. Right row5,col6 Y present. So shape is like a T? Actually plus with extra fin left or right? The description: "plus-shaped post with one extra fin jutting to the left or right". So base plus shape (center plus up/down/left/right) plus an extra fin either left or right of the center? In example1, extra fin is at row4,col6 (above right). That's extra fin to the right? Hmm.

Weather vane Y pattern: In both examples, Y occupies middle three columns rows 4-6. Let's see Example2 input Y pattern same as example1? Example2 rows 4-6 same as example1: row4 . . . Y . . . . . => Y at col4? Wait row4: ". . . Y . . . . ." => Y at col4. Row5: ". . . Y Y Y . . ." => Y at col4,5,6. Row6: ". . . . Y . . . ." => Y at col5. So here extra fin is at row4,col4 (above left). So extra fin left side.

Thus the weather vane orientation indicates direction: extra fin left => wind direction left? Or right? The description: plus-shaped post with one extra fin jutting to the left or right, revealing which way the wind is blowing. So extra fin indicates direction of wind. In example1 extra fin at right (col6) => wind blowing right? So cloud reflection should be placed to the right of original cloud. In output example1, we see additional A's added to the right side of original cloud.

Let's compare input vs output example1.

Input rows 1-3 have A cloud shape. Output rows 1-3 have more A's.

Row1 input: A at col4 and col6. Output row1: ". . . A . A A . A". Let's list columns: col1 .,2 .,3 .,4 A,5 .,6 A,7 A,8 .,9 A? Wait there are 9 columns. Output row1: ". . . A . A A . A". That's columns: 1 .,2 .,3 .,4 A,5 .,6 A,7 A,8 .,9 A. So new A at col7 and col9. Original had at col4 and col6. So added A's to the right side (col7,9). Row2 input: A at col4,5,6. Output row2: ". . . A A A A A A". That's A at col4-9 all A except maybe col8? Actually 

... [10127 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 444: "A seed crystal grows into ever-larger reflections of itself, ... [35 chars truncated]" k=1 alts: 8/15 762cd429

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black background with a single 2x2 colored seed pattern placed somewhere in the grid. The seed contains four colored cells arranged in a 2x2 block (top-left, top-right, bottom-left, bottom-right), where each cell is a non-black color. The seed's four colors define a quadrant pattern.

Do: Starting from the 2x2 seed pattern, generate a series of scaled copies at exponentially increasing sizes, each placed immediately to the right of the previous one and vertically centered on the seed's position. Specifically: Level 0 is the original 2x2 seed (each cell is 1x1). Level 1 doubles the scale: each of the four quadrant colors becomes a 2x2 block, producing a 4x4 region placed directly to the right of level 0 and centered vertically on the seed. Level 2 doubles again: each quadrant color becomes a 4x4 block, producing an 8x8 region placed to the right of level 1, centered on the seed. Continue doubling (level n has cell size 2^n, producing a region of size 2^(n+1) x 2^(n+1)) until the grid's right edge is reached. At every level, the 2x2 quadrant arrangement is preserved: top-left quadrant uses the seed's top-left color, top-right uses the seed's top-right color, bottom-left uses the seed's bottom-left color, bottom-right uses the seed's bottom-right color. Each level's region is vertically centered on the seed's center row. Cells that fall outside the grid boundaries remain black.

Grid: The output grid has the same dimensions as the input grid. The background is black. The seed pattern remains at its original position. The scaled copies expand rightward from the seed, filling the grid toward the right edge. The vertical extent of each level is centered on the seed's vertical midpoint, growing symmetrically upward and downward. No content extends to the left of the seed.

Figurative Descriptions (16 variants)

"A seed crystal grows into ever-larger reflections of itself, each generation dwarfing the last."

120b: k=1

See: In a vast dark field, a tiny 2×2 embryo sits quietly — four colored organs packed into one small heart. Each quadrant holds a different hue, like the four chambers of a beating heart waiting to pump life outward.

Do: The heart begins to beat. With each pulse, it projects a shadow of itself to the right — but each shadow is twice the size of the last, as if viewed through a succession of magnifying glasses lined up in a row. The first echo is a 4×4 ghost, the next an 8×8 phantom, then 16×16, and so on, each one faithfully preserving the four-chambered arrangement of the original heart, just writ larger and larger. The echoes march rightward like Russian nesting dolls laid open in a line, smallest to largest, until they crash against the edge of the world. Each echo is vertically anchored to the heartbeat's center, growing equally above and below like a tree branching symmetrically from its trunk.

Grid: The dark field stays the same size. The original heart never moves. Its echoes only radiate to the right, never to the left. Where an echo would extend beyond the sky or below the earth, it simply doesn't exist — the darkness holds. The four colors in each echo always sit in the same quadrant positions as the original: upper-left stays upper-left, lower-right stays lower-right, like a flag whose design never rotates no matter how large the banner grows.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (12194 chars)

We need to understand rule. From description: original 2x2 embryo somewhere in grid. It stays same. Echoes to the right, each larger by factor 2 each step (size doubles). They are anchored vertically centered on original's center, expanding equally up/down. They are placed to the right, not overlapping? In example 1, input shows a 2x2 block at rows 5-6? Let's map coordinates. Input grid size? appears 10 rows, 14 columns? Actually given lines: first row ". . . . . . . . . . . . . ." (14). So 10 rows, 14 columns.

Input has R X at (row5 col1?) Let's index rows starting 1. Row5: "R X . . . . . . . . . . . ." So R at col1, X at col2. Row6: "X G . . . . . . . . . . . ." So X at col1, G at col2. That's a 2x2 block top-left corner of grid.

Output: same size grid 10x14. Let's see output rows.

Row1 all dots.

Row2: ". . . . . . R R R R X X X X" So starting at col7? Let's count: columns 1-6 dots, then R repeated 4, X repeated 4. So a 4x4 echo placed starting column7? Row2 corresponds to top part of 4x4 block? The original 2x2 block is at rows5-6, cols1-2. Its center is between rows5-6? Actually 2x2 center is between those cells. For vertical anchoring, the 4x4 block should be centered on same vertical center, meaning its rows should be rows4-7? Since original rows5-6, center between 5 and6. A 4x4 block centered would occupy rows4-7 (since height 4). Indeed output rows4-7 contain pattern.

Check output rows4-7 columns? Row4: ". . R R X X R R R R X X X X". That's columns 3-? Let's count: col1-2 dots, then R R X X (first 4 columns of block?), then R R R R X X X X (second 4?). Wait pattern: The 4x4 block should be composed of 2x2 pattern scaled by 2: each original cell becomes 2x2 block of same color. Original arrangement:
R X
X G

Scaled 2x2 each => 4x4:
R R X X
R R X X
X X G G
X X G G

Now output shows something like that but also repeated? Actually they placed two copies? Let's examine output rows2-3: they have R R R R X X X X (8 columns). That's like two 4x4 blocks side by side? Wait row2: after 6 dots, there are 4 R then 4 X. That's 8 columns. Row3 same. Row4: ". . R R X X R R R R X X X X". That's 2 dots then 2 R,2 X, then 4 R,4 X? Not matching.

Maybe the echo pattern repeats horizontally as it expands: each echo is twice size of previous, but placed to the right, not overlapping, so they are tiled horizontally: first echo 4x4 placed immediately right of original (starting col3?). But output shows first echo starting at col3? Actually original occupies col1-2. 4x4 echo would occupy cols3-6. But output rows4-7 columns3-6 show pattern? Let's check row4 col3-6: "R R X X" matches top row of scaled block. Row5 col3-6: "R R X X"? Row5 line: "R X R R X X R R R R X X X X". Let's parse row5: col1 R, col2 X (original), col3 space? Actually row5: "R X R R X X R R R R X X X X". So after original two cells, we have "R R X X" (col3-6) which matches top row of block. Row6: "X G X X G G X X X X G G G G". After original "X G", we have "X X G G". That's second row of

... [9215 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 449: "A dam holds back a flood that fills everything below it." k=2 alts: 12/15 782b5218

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains cells of three types: red (R) cells that form a diagonal or irregular boundary/separator line, a dominant non-red color (blue, grey, azure, or maroon depending on the puzzle instance) scattered throughout the grid on both sides of the red boundary, and black/empty (.) cells also scattered on both sides. The red cells divide each column into an 'above' region and a 'below' region.

Do: 1. Identify all red (R) cells in the grid. These remain fixed in place throughout the transformation. 2. Identify the non-red, non-black color present in the grid (e.g., blue, grey, azure, maroon) — this is the fill color. 3. For each column independently, find the red cell(s). Every non-red cell ABOVE the topmost red cell in that column becomes black (.). Every non-red cell BELOW the bottommost red cell in that column becomes the fill color. 4. Red cells stay exactly where they are. The result is that the region above the red boundary is cleared to empty/black, the red boundary is preserved, and the region below the red boundary is filled solid with the fill color.

Grid: The grid is square (10x10 in all examples). It uses a black (.) background with exactly two other colors: red (R) serving as a fixed boundary, and one other color serving as fill material. The red cells form a connected or near-connected band that spans the full width of the grid, separating it into an upper and lower region.

Figurative Descriptions (16 variants)

"A dam holds back a flood that fills everything below it."

120b: k=2

See: A red dam stretches across the landscape, sometimes straight, sometimes staggered like a staircase. Above and below the dam, colored sediment and empty air pockets are scattered randomly — the remnants of a world before the flood.

Do: When the floodwaters rise, the dam holds firm. Everything above the dam is swept clean — scoured to empty sky. Everything below the dam is submerged: the water fills every pocket, every gap, drowning the space in a uniform layer of sediment. The dam itself, red and unyielding, stays exactly where it was built.

Grid: The world is a 10×10 valley. The red dam cuts across it column by column, and the colored sediment is the same substance everywhere — it just needs the flood to press it into place. Above: clear sky. Below: solid earth. The dam: immovable.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (11958 chars)

We need to understand the puzzle. The description: red dam across landscape, sometimes straight, sometimes staggered like a staircase. Above dam: clear sky (empty). Below dam: solid earth (filled). The dam stays same. So transformation: For each column, there is a red row (R) that forms a dam across columns, maybe not same row each column? In examples, we see input grid 10x10. In example1, there is a row of R's across entire row 4 (0-index? Actually row index 3 if starting 0). Input shows row 4 (line 4) is "R R R R R R R R R R". That's the dam. Above rows have various colors (B, .). Below rows have B and . etc. Output: rows above dam become all '.' (black). Rows below dam become all B (blue) (the color of sediment). The dam row stays R.

But note: In example1, below dam rows originally had mixture of B and .; after transformation they become all B. So the sediment color is B (blue). The rule: fill all cells below the dam with the most common non-R color? Actually in example1, the only other color present is B (blue). So fill with B.

In example2, input has a row of R's but not a full row? Let's examine example2 input: lines:

1: X X X . X . . . X X
2: X . . X X . X . X X
3: . X X . X X . X . .
4: R . X X R . X . R X
5: X R . R . R . R . R
6: . . R X X X R . X .
7: X X . . . X X X X X
8: . X . X X . X . X X
9: . X X . X . X . X X
10: X X . . X X X X X X

We see a "dam" of R's but not a straight line: row4 has R at columns 1,5,9? Actually row4: "R . X X R . X . R X" => R at col1, col5, col9 (1-index). Row5: "X R . R . R . R . R" => R at col2,4,6,8,10. Row6: ". . R X X X R . X ." => R at col3,7. So the dam is like a staircase shape: a contiguous path of R's that goes downwards? The description says dam cuts across column by column. So each column has exactly one R cell? Let's check columns: column1: rows: X, X, ., R, X, ., X, ., ., X => R at row4 only. column2: X, ., X, ., R, ., X, X, X, X => R at row5 only. column3: X, ., X, X, ., R, ., ., X, . => R at row6 only. column4: ., X, ., X, R, X, ., X, ., . => R at row5? Wait column4 row5 is R, row4 is X, row6 is X. So R at row5. column5: X, X, X, R, ., X, ., X, X, X => R at row4. column6: ., ., X, ., R, X, X, ., ., X => R at row5. column7: ., X, ., X, ., R, X, X, X, . => R at row6. column8: ., ., X, ., R, ., X, ., ., X => R at row5. column9: X, X, ., R, ., R, X, X, X, . => R at row4 and row6? Actually row4 col9 is R, row6 col9 is R. That's two R's in same column, which violates one per column. Let's double-check: row6 is ". . R X X X R . X ." => col9 is X, not R. So only row4 col9 is R. column10: X, X, ., X, R, ., X, X, X, X => R at row5.

Thus each column has exactly one R, forming a diagonal staircase from top-left to bottom-right? Let's list rows of R per column (1-indexed):
col1: row4
col2: row5
col3: row6
col4: row5
col5: row4
col6: row5
col7: row6
col8: row5
col9: row4
col10: row5

So pattern: rows vary but each column has exactly one R. The dam is a set of

... [9078 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.02)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 454: "A compass rose unfurls from a single petal, each cardinal di ... [47 chars truncated]" k=2 alts: 15/15 7953d61e

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: A 4×4 grid of colored cells. The grid is treated as a single rigid tile that can be rotated.

Do: Create an 8×8 output by tiling four rotated copies of the 4×4 input into a 2×2 arrangement of quadrants: (1) Top-left quadrant: the original input, unchanged. (2) Top-right quadrant: the input rotated 90° counter-clockwise (equivalently, transpose then reverse each column). (3) Bottom-left quadrant: the input rotated 180°. (4) Bottom-right quadrant: the input rotated 90° clockwise (equivalently, transpose then reverse each row).

Grid: The input is always a 4×4 grid. The output is always an 8×8 grid composed of four 4×4 quadrants arranged in a 2×2 layout. There is no background color or border; every cell is determined by the corresponding rotation of the input.

Figurative Descriptions (16 variants)

"A compass rose unfurls from a single petal, each cardinal direction revealing a new face of the same bloom."

120b: k=2

See: A small mosaic tile — a 4×4 swatch of stained glass — sits at the heart of a cathedral floor, waiting to be echoed outward.

Do: The tile dreams of itself from every direction. It gazes north and sees its own face. It turns its head east, west, and south, and at each glance the reflection has spun — as if the tile were placed on a lazy Susan and photographed at each quarter-turn. These four portraits are pressed together like pages of a folded map, the original anchoring one corner while its dizzied twins fill the remaining three.

Grid: The cathedral floor is exactly four tiles wide and four tiles tall — a 2×2 arrangement of the original swatch and its three rotated echoes. No grout, no gaps: every cell of colored glass belongs to one of the four spinning memories of the original piece.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (7466 chars)

We need to infer rule. Input is 4x4, output is 8x8 (since examples output 8x8). The description: original 4x4 tile placed at heart, then rotated copies (90°,180°,270°) placed to form a 2x2 arrangement. Likely output is a 8x8 where quadrants are the original and its rotations. Need to determine arrangement: which quadrant gets which rotation? The description: "original anchoring one corner while its dizzied twins fill the remaining three." So original placed at one corner of 8x8 (maybe top-left) and other three quadrants are rotated versions.

Check Example1. Input 4x4 grid:

Row1: Y W B A
Row2: A Y B A
Row3: Y A A B
Row4: B B B A

Let's compute rotations.

Original (0°):
[ [Y,W,B,A],
  [A,Y,B,A],
  [Y,A,A,B],
  [B,B,B,A] ]

Rotate 90° clockwise: new[i][j] = original[3-j][i] (size 4). Let's compute.

Indices: rows 0-3, cols 0-3.

Compute rotated90:
row0: original[3][0]=B, [2][0]=Y, [1][0]=A, [0][0]=Y => B Y A Y
row1: original[3][1]=B, [2][1]=A, [1][1]=Y, [0][1]=W => B A Y W
row2: original[3][2]=B, [2][2]=A, [1][2]=B, [0][2]=B => B A B B
row3: original[3][3]=A, [2][3]=B, [1][3]=A, [0][3]=A => A B A A

Rotate 180°: rotate 90 twice or reverse both axes. Let's compute: row0 = original[3][3], original[3][2], original[3][1], original[3][0] => A B B B
row1: original[2][3], [2][2], [2][1], [2][0] => B A A Y
row2: original[1][3], [1][2], [1][1], [1][0] => A B Y A
row3: original[0][3], [0][2], [0][1], [0][0] => A B W Y

Rotate 270° clockwise (or 90° ccw): row0 = original[0][3], [1][3], [2][3], [3][3] => A A B A
row1 = original[0][2], [1][2], [2][2], [3][2] => B B A B
row2 = original[0][1], [1][1], [2][1], [3][1] => W Y A B
row3 = original[0][0], [1][0], [2][0], [3][0] => Y A Y B

Now output 8x8 given. Let's layout output rows:

Row1: Y W B A A A B A
Row2: A Y B A B B A B
Row3: Y A A B W Y A B
Row4: B B B A Y A Y B
Row5: A B B B B Y A Y
Row6: B A A Y B A Y W
Row7: A B Y A B A B B
Row8: A B W Y A B A A

We need to see quadrants 4x4 each.

Top-left quadrant (rows1-4, cols1-4) = original? Let's extract:

Rows1-4, Col1-4:
Row1: Y W B A
Row2: A Y B A
Row3: Y A A B
Row4: B B B A

Yes matches original. Good.

Top-right quadrant (rows1-4, cols5-8):
Rows1: A A B A
Row2: B B A B
Row3: W Y A B
Row4: Y A Y B

That appears to be rotation? Compare with 90°? Our 90° rotation matrix was:
B Y A Y
B A Y W
B A B B
A B A A

Not match. 180° rotation matrix:
A B B B
B A A Y
A B Y A
A B W Y

Not match. 270° rotation matrix:
A A B A
B B A B
W Y A B
Y A Y B

Yes top-right matches 270° rotation (clockwise). Good.

Bottom-left quadrant (rows5-8, cols1-4):
Rows5: A B B B
Row6: B A A Y
Row7: A B Y A
Row8: A B W Y

That matches 180° rotation? Let's compare 180° matrix we computed:
A B B B
B A A Y
A B Y A
A B W Y

Exactly matches. Good.

Bottom-right quadrant (rows5-8, cols5-8):
Rows5: B Y A Y
Row6: B A Y W
Row7: B A B B
Row8: A B A A

That matches 90° rotation matrix? 90

... [4882 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.06)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 460: "A stained-glass window reveals its true colors when sunlight ... [24 chars truncated]" k=2 alts: 9/15 7c008303

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a 9×9 grid divided into four quadrants by a horizontal row and a vertical column of azure (A) cells that form a cross. One of the four quadrants (6×6) contains a pattern made of green (G) cells on a black (.) background. An adjacent quadrant (2×2, tucked into the corner next to the A-cross) contains exactly four colored cells arranged in a 2×2 block — these are the color key. The remaining two quadrants are empty (all black).

Do: 1. Locate the row and column of A cells; they intersect to divide the grid into four quadrants. 2. Identify the 6×6 quadrant containing the green pattern and the 2×2 quadrant containing the color key. 3. Determine the spatial relationship between the key quadrant and the pattern quadrant: the 2×2 key maps onto the four 3×3 sub-blocks of the 6×6 pattern based on their relative positions (top-left key cell maps to the top-left 3×3 block, top-right key cell to the top-right 3×3 block, bottom-left key cell to the bottom-left 3×3 block, bottom-right key cell to the bottom-right 3×3 block). 4. Produce a 6×6 output grid: for each 3×3 block of the green pattern, replace every green (G) cell with the color from the corresponding position in the 2×2 key, and keep every black (.) cell as black.

Grid: The input grid is always 9×9. The output grid is always 6×6 — it contains only the recolored pattern with no azure dividers. The background color is black (.). The 2×2 color key can appear in any of the three non-pattern quadrants (its exact corner varies across instances), and the green pattern can occupy any of the four quadrants, but it is always the 6×6 one.

Figurative Descriptions (16 variants)

"A stained-glass window reveals its true colors when sunlight passes through a prism."

120b: k=2

See: An azure cross divides the world like the leading of a cathedral window, separating four panes. One large pane holds a pattern traced in plain green glass — a design waiting to be illuminated. Tucked into a nearby corner, a tiny 2×2 prism glows with four jewel-toned colors, each one a secret instruction. The remaining panes are dark, as if the light hasn't reached them yet.

Do: The prism catches the light and casts it onto the green glass. Each of the prism's four facets shines toward the quadrant of the pattern that mirrors its own position — upper-left facet illuminates the upper-left quarter, and so on. Wherever the green glass stands, it absorbs the color of the light falling on it, transforming from plain green into the jewel tone of its corresponding facet. Wherever there is only darkness, no light takes hold. Strip away the leading — the azure framework was only ever scaffolding — and what remains is the stained-glass window in full color.

Grid: The cathedral window is always 9×9 before the leading is removed. The finished stained-glass panel is always 6×6 — the illuminated design without its frame. The darkness is black (.). The prism may sit in any dark corner adjacent to the leading, and the green tracery may fill any of the four panes, but it always occupies the large one.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (11977 chars)

We need to understand puzzle. Input grid 9x9 with azure leading cross? In examples, there is a column and row of A (azure) forming cross dividing into four quadrants. The large pane contains green pattern (G). The small 2x2 prism of four colors (different from green and azure) located in a corner adjacent to the leading (i.e., in one of the quadrants, at corner). The prism is 2x2 block of distinct colors (e.g., in example1: top-left quadrant? Let's examine.

Example1 Input grid (9x9). Let's layout rows:

Row1: . . . . . . A B R
Row2: . . . . . . A Y B
Row3: A A A A A A A A A
Row4: . . G G . G A . .
Row5: G G . . . . A . .
Row6: G G . G . G A . .
Row7: . . . . G . A . .
Row8: G G G G G G A . .
Row9: . . . . G . A . .

We see column 7 (index 7?) Actually A's column at col7? Let's index 1-9. Row1 col7 = A, col8 = B, col9 = R. Row2 col7 = A, col8 = Y, col9 = B. Row3 all A across. So cross: vertical line of A at column7 (col7) and horizontal line of A at row3 (row3). That divides into four quadrants: top-left (rows1-2, cols1-6), top-right (rows1-2, cols8-9), bottom-left (rows4-9, cols1-6), bottom-right (rows4-9, cols8-9). The large pane holds green pattern: appears in bottom-left quadrant (rows4-9, cols1-6) there are G's forming shape. Indeed.

The prism is a 2x2 block of colors in top-right quadrant? Let's see top-right quadrant rows1-2, cols8-9: we have (row1 col8 B, col9 R; row2 col8 Y, col9 B). That's a 2x2 block with colors B,R,Y,B (not all distinct). But description says four jewel-toned colors each one a secret instruction. In example1, the colors are B,R,Y,B (B repeats). Hmm maybe the prism is not necessarily all distinct; but each facet corresponds to quadrant: upper-left facet illuminates upper-left quarter, etc. So the prism's positions correspond to quadrants: top-left facet (row1 col8?) Wait orientation: The prism sits in a corner adjacent to the leading. In example1, the prism is in top-right corner (adjacent to both the vertical and horizontal A lines). The 2x2 block occupies rows1-2, cols8-9. Its four cells correspond to quadrants: upper-left cell of prism (row1 col8) corresponds to upper-left quadrant (top-left). Upper-right cell (row1 col9) corresponds to upper-right quadrant. Lower-left cell (row2 col8) corresponds to lower-left quadrant. Lower-right cell (row2 col9) corresponds to lower-right quadrant. That matches "upper-left facet illuminates the upper-left quarter, and so on."

Now transformation: For each quadrant, wherever there is green G in the large pane, replace it with the color of the corresponding prism cell. If there is no green (i.e., black .) remain black. Also maybe other colors (non-green) become something? In output example1, we see output 6x6 (after removing leading cross). Let's compute.

First, we need to extract the 6x6 region that is the large pane (the quadrant with green). In example1, large pane is bottom-left quadrant (rows4-9, cols1-6). That's 6 rows (4-9) and 6 columns (1-6). Indeed output is 6x6.

Now 

... [9034 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.14)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 461: "Stained glass windows are filled by the dominant color of th ... [27 chars truncated]" k=1 alts: 2/15 7c8af763

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The 10x10 grid is partitioned into rectangular regions by horizontal and vertical separator lines. Separator lines are composed of grey (X) cells interspersed with colored markers (R or B). Horizontal separator lines span entire rows; vertical separator lines span entire columns. Together they form a lattice that divides the grid into rectangular compartments. Inside each compartment, the cells are empty (black dots). The colored markers (R, B) embedded in the separator lines serve as boundary labels for the adjacent compartments.

Do: 1. Identify all horizontal separator rows and vertical separator columns. These are the rows and columns composed entirely of X, R, and B (no dot cells). They form a grid-like lattice. 2. The intersections of these separator lines define rectangular interior regions (compartments) whose cells are all dots in the input. 3. For each compartment, collect all colored (non-X) markers that lie on its four boundary edges — that is, the colored markers on the separator row above, the separator row below, the separator column to the left, and the separator column to the right, restricted to the span of that compartment's extent. Include corner cells shared with adjacent separators. 4. Determine the majority color among these collected boundary markers (whichever of R or B appears more often). 5. Replace every dot cell inside that compartment with the majority color. 6. All separator cells (X and colored markers) remain unchanged in the output.

Grid: The grid is always 10x10. The background fill is black (dots) inside compartments. Grey (X) cells and colored markers (R, B) form separator lines that persist unchanged through the transformation. Separator lines run the full width or height of the grid, creating a lattice of rectangular compartments. Each compartment is at least 1 cell wide and 1 cell tall. The only colors used are black (.), grey (X), red (R), and blue (B).

Figurative Descriptions (16 variants)

"Stained glass windows are filled by the dominant color of their surrounding lead frame."

120b: k=1

See: A cathedral wall of grey (X) lead caming forms a lattice of window panes. Embedded in the lead are gemstones — rubies (R) and sapphires (B) — set into the framework like jeweled accents. The window panes themselves are empty (black), waiting to be filled with colored glass.

Do: Each empty pane yearns to be filled with light. It looks to the jewels set into its surrounding frame — the rubies and sapphires adorning its four edges — and takes on the hue of whichever gem appears more often. Where more rubies border a pane, it floods with red light; where sapphires dominate, it glows blue. The lead caming and its embedded jewels remain untouched — only the empty glass transforms.

Grid: The cathedral wall is always 10×10 stones. Grey lead (X) and colored jewels (R, B) form the rigid skeleton that never changes. The black (.) panes sit within rectangular openings in this lattice, each pane at least one stone wide and tall. Only four materials exist: lead (X), ruby (R), sapphire (B), and empty darkness (.) waiting for its color.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (11741 chars)

We need to infer rule: For each '.' cell, look at its four orthogonal neighbors (up, down, left, right). Count R and B among those neighbors (maybe also X? but description says rubies and sapphires on four edges). Choose whichever appears more often; if tie? maybe keep '.'? In example, we can test.

Let's reconstruct example 1 input and output to verify.

I'll map coordinates 10x10.

Input grid (I'll index rows 0-9, cols 0-9). Represent.

Row0: R X X X B X X X X R
Row1: . . X . . . . X . .
Row2: . . B . . . . B . .
Row3: . . X . . . . X . .
Row4: . . X . . . . X . .
Row5: X B X X X R X X X B
Row6: . . X . . . . X . .
Row7: . . X . . . . R . .
Row8: . . X . . . . X . .
Row9: B X X X R X X X X B

Output grid:

Row0 same.
Row1: B B X B B B B X B B
Row2: B B B B B B B B B B
Row3: B B X B B B B X B B
Row4: B B X B B B B X B B
Row5 same as input row5.
Row6: B B X R R R R X B B
Row7: B B X R R R R R B B
Row8: B B X R R R R X B B
Row9: B X X X R X X X X B

We need to see transformation of '.' cells only. Let's compute for a sample.

Take row1 col0 (input '.'), neighbors: up row0 col0 = R, down row2 col0 = '.' (but that's also '.' originally), left none, right col1 '.' . So only R neighbor? Also maybe consider only non '.'? The rule says look at rubies and sapphires on four edges. So count R and B among the four orthogonal cells, ignoring '.' and X? In this case up is R, down is '.' (ignore), left none, right '.' (ignore). So R count=1, B count=0 => choose R? But output at row1 col0 is B (blue). Hmm not matching.

Maybe they consider the four cells that are part of the "frame": the immediate orthogonal cells that are not '.' (i.e., X, R, B). In this case down is '.' but maybe later after transformation? Not.

Let's examine row1 col0 output B. So maybe rule is opposite: choose the majority among the four cells (including X?) but X is grey not counted? Let's count: up R, down '.' (maybe treat as empty), left none, right '.' . So R majority, would be R not B.

Thus maybe we misinterpret coordinates. Let's double-check input row1 col0 is '.'? Row1: ". . X . . . . X . ." yes first two are '.'.

But output row1 col0 is B. So something else.

Maybe the rule is to fill each '.' region (connected component) with the color that is most common among the border cells (R or B) surrounding the region (i.e., the region's perimeter). The region may be larger than single cell. In example, there are rectangular openings (windows) comprised of '.' cells forming rectangles. The rule: each empty pane (i.e., each rectangular region) takes color of majority of gems on its four edges (the border of the rectangle). That matches description: "Each empty pane yearns to be filled ... looks to the jewels set into its surrounding frame — the rubies and sapphires adorning its four edges — and takes on the hue of whichever gem appears more often."

Thus we need to identify each rectangular region of '.' cells (connected orthogonally) that is bounded by X, R, B (the frame)

... [8813 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.01)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 467: "A comet streaks away from its home planet, leaving a glowing ... [28 chars truncated]" k=1 alts: 13/15 7ddcd7ec

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains a single 2x2 solid block of a colored cell (non-black) and one or more isolated single cells of the same color positioned diagonally adjacent to corners of the 2x2 block. These single cells act as diagonal arms indicating ray directions extending outward from the block.

Do: Keep the original 2x2 block and all original single cells in place. For each single cell that is diagonally adjacent to a corner of the 2x2 block, identify the diagonal direction pointing away from the block (i.e., the direction from the block's corner to that single cell). Starting from that single cell, continue placing one cell of the same color per row along that diagonal, stepping one row and one column in the same direction each time, until reaching the edge of the grid.

Grid: The grid is 10x10 with a black (.) background. All colored cells share the same single color. The 2x2 block is positioned away from the edges, leaving room for diagonal rays to extend to the grid boundary.

Figurative Descriptions (16 variants)

"A comet streaks away from its home planet, leaving a glowing trail across the night sky."

120b: k=1

See: In the darkness of space, a small colored planet floats — a solid 2×2 world. Nearby, one or two comets have just launched from its corners, each a single bright spark sitting diagonally adjacent to the planet's surface.

Do: Each comet is hurtling away from the planet. It blazes a trail across the void, dropping one spark of light in every row it crosses, always flying diagonally outward until it crashes into the edge of the universe. The planet itself stays fixed — only the comets move, and they never look back.

Grid: The universe is a 10×10 field of black void. The planet and its comets share one color — they belong to the same star system. The planet drifts far enough from the edges that each comet has runway to streak across the sky before hitting the boundary.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (5939 chars)

We need to infer rule. From description: planet is 2x2 block, comets are single cells diagonally adjacent to planet corners. They move outward diagonally, leaving a trail: one spark per row they cross, until hitting edge. The planet stays fixed. So input has a 2x2 block of a color (same as comets). The comets are at positions diagonally adjacent to the block corners (i.e., at (row-1,col-1) etc). They then move outward: each step they go one row and one column further away (same direction) until edge, and they fill cells along the way? "dropping one spark of light in every row it crosses" meaning they fill each cell along its diagonal path? In example 1, color Y. Let's examine input and output.

Example1 Input grid: Let's map coordinates (0-index). We'll list Y positions.

Rows:
0: none
1: column6? Actually row1: ". . . . . . Y . . ." => column6 (0-index) = Y.
2: ". . . . Y Y . . . ." => columns4,5 Y.
3: same as row2 columns4,5 Y.
4: ". . . . . . Y . . ." => column6 Y.
others none.

So shape: a 2x2 block at rows2-3, cols4-5 (that's the planet). Comets: there are Y at (1,6) which is diagonally adjacent to top-right corner? Planet top-right corner is (2,5). Diagonal up-right from that is (1,6) yes. Also Y at (4,6) which is diagonal down-right from bottom-right corner (3,5) -> (4,6). So there are two comets on the right side only. No left side comets.

Output Y positions:
Row0: column7? ". . . . . . . Y . ." => col7 Y.
Row1: column6 Y.
Row2: cols4,5 Y (same planet)
Row3: cols4,5 Y
Row4: column6 Y (same as input)
Row5: column7 Y.
Row6: column8 Y.
Row7: column9 Y.
Rows8-9 none.

So the comet at (1,6) moved outward up-right direction: from (1,6) to (0,7) then hits edge? Actually continues until column9? Wait up-right from (1,6) goes to (0,7) then next would be (-1,8) out of grid, so stops at (0,7). That's present in output row0 col7 Y. Good.

The comet at (4,6) moves down-right: (5,7), (6,8), (7,9) then out. Output includes Y at (5,7), (6,8), (7,9). Yes.

Thus rule: For each comet cell that is diagonally adjacent to the 2x2 block, extend a line of same color outward in same diagonal direction until hitting grid edge, filling each cell along the way (including original comet? original stays? In output, original comet positions remain? Yes (1,6) and (4,6) still present). So we keep original comet and add trail cells beyond it outward.

Now test input: we need to locate 2x2 block of same color A (azure). Input grid:

Rows:
0 all .
1 all .
2 all .
3: ". . . . . A . . . ." => column5? Actually count: indices 0-9. Row3 col5 = A.
4: ". . . A A . . . . ." => col3,4 = A A.
5: same row5 col3,4 = A A.
6: ". . A . . A . . . ." => col2 = A, col5 = A.
7-9 all .

Let's map positions.

Planet: 2x2 block likely at rows4-5, cols3-4 (since that's a 2x2 of A). Yes.

Now comets: cells diagonally adjacent to any corner of that block.

Corners of block:
Top-left: (4,3)
Top-right: (4,4)
Bottom-left: (5,3)
Bottom-right: (5,4)

Check diagonal adjacent cel

... [2969 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.98)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 488: "A beacon sinks one step into the earth, and its light rains ... [69 chars truncated]" k=1 alts: 2/15 834ec97d

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black background with exactly one non-black colored cell located somewhere in the grid.

Do: 1. Locate the single colored cell at row R, column C. Note its color. 2. Determine the parity (even or odd) of column C. 3. Move the colored cell down by one row to position (R+1, C). 4. Fill every cell from row 0 through row R (inclusive) with yellow, but ONLY at columns that share the same parity as C. All other cells in those rows remain black. 5. All rows below the colored cell's new position remain unchanged (black).

Grid: The grid dimensions match the input. The background is black (.). The coordinate system is zero-indexed with row 0 at the top and column 0 at the left. The grid contains exactly one non-black cell in the input.

Figurative Descriptions (16 variants)

"A beacon sinks one step into the earth, and its light rains upward in alternating columns like prison bars against the night sky."

120b: k=1

See: A single beacon glows alone in a dark field — a solitary flame in an ocean of night.

Do: The beacon is heavy; it sinks one step deeper into the ground. As it falls, it casts its light upward like rain falling in reverse, illuminating every other column above where it once stood — golden bars of light striping the darkness. The columns that catch the light share a kinship with the beacon's own column, stepping in rhythm with it (every other one). The sky above fills with these golden stripes, but the darkness below the beacon remains untouched, as light cannot fall downward from a thing that has already fallen.

Grid: The dark field keeps its original shape. Night is black (.), and the beacon's golden rain is yellow (Y). The beacon itself keeps its original color and identity. Only the sky above the beacon's resting place is painted — the ground beneath stays dark and silent.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (3624 chars)

We need to infer rule from example. Input grid 5x5 with a single M at (row2,col2) zero-index? Let's map coordinates: rows 0-4, columns 0-4. Input shows M at row2 col2 (center). Output: Y . Y . Y in rows 0-2? Actually output grid 5x5: rows 0-2 are Y . Y . Y, row3 is . . M . ., row4 all . . . . . So rule: beacon (M) "sinks one step deeper" meaning moves down one row? In output, M stays at same position? Wait input M at row2 col2, output M at row3 col2 (one row down). Yes M moved down one row. The "golden rain" Y appears in columns that are every other column (alternating) above the new position? Let's see: Y appears in columns 0,2,4 (every other column starting at 0) for rows 0-2 (above the new M row). So Y fills all cells above the new M row, but only in columns that are same parity as M column? M column is 2 (even). The Y columns are 0,2,4 (even columns). So rule: after moving beacon down one row, fill all cells above it (i.e., rows less than new row) in columns that have same parity (even/odd) as beacon's column with Y. The rest remain unchanged (original colors). Also note that original M's original location becomes . (black) because M moved down, leaving empty. Indeed row2 col2 becomes Y? Wait row2 col2 in output is Y (since row2 is above new M row (row3) and column 2 matches parity). So original M location becomes Y, not . So the original cell is overwritten by Y if column matches parity. So rule: shift the special element (beacon) down one row (if possible). Then for all cells above the new position (i.e., rows 0 to newRow-1), set Y in columns where (col - beaconCol) mod 2 == 0 (i.e., same parity). Keep other cells unchanged (including other colors). In example, there were only black cells elsewhere.

Now test input: 12x12 grid with a single G at row3 col5? Let's count: rows 0-11, columns 0-11. Input shows G at row3 (0-index) column5? Let's verify: line: ". . . . . G . . . . . ." That's row3, columns: 0 .,1 .,2 .,3 .,4 .,5 G,6 ., etc. So beacon color is G (green). According to rule, beacon moves down one row to row4 same column5. Then we fill Y in columns with same parity as column5 (odd columns) for rows above new row (rows 0-3). Column5 is odd (5). So Y in columns 1,3,5,7,9,11? Wait columns 0-11, odd columns are 1,3,5,7,9,11. So for rows 0-3, set Y at those columns. However need to consider that original G cell at row3 col5 will be overwritten by Y (since row3 is above new row4). So final grid: rows 0-3 have Y at odd columns, other cells remain . (black). Row4 will have G at col5 (the moved beacon). Row5-11 remain unchanged (all .). Also need to keep any other colors? None.

Edge case: what if moving down goes out of bounds? Not here.

Thus produce output 12x12.

Let's construct rows.

Row0: columns 0-11: col0 ., col1 Y, col2 ., col3 Y, col4 ., col5 Y, col6 ., col7 Y, col8 ., col9 Y, col10 ., col11 Y. So pattern . Y . Y . Y . Y . Y . Y

Row1 same as row0.

Row2 same.

Row3 same (since still above new row). Row3 originally had G at col5, bu

... [624 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.67)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 496: "A walled city where open ground exposed to the wilderness be ... [78 chars truncated]" k=2 alts: 3/15 84db8fc4

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains three types of cells: G (green), B (blue), and . (black/empty). The G and B cells form a fixed structural framework, while the black cells fill the remaining spaces. The black cells naturally form connected components — groups of adjacent black cells linked by orthogonal (up/down/left/right) adjacency. Some of these connected components touch the border of the grid (i.e., at least one cell sits on the first or last row, or the first or last column), while others are entirely enclosed within the interior, surrounded on all sides by G and B cells.

Do: 1. Identify all cells containing G or B. These remain unchanged in the output. 2. Identify all black (.) cells. 3. Find the connected components of black cells using 4-connected (orthogonal) adjacency — two black cells belong to the same component if they share an edge horizontally or vertically. 4. For each connected component, determine whether any cell in the component lies on the grid border (row 0, last row, column 0, or last column). 5. If the component touches the border, replace every black cell in that component with R (red). 6. If the component is entirely interior (no cell on any border edge), replace every black cell in that component with X (grey).

Grid: The grid is a fixed 10×10 square. Rows are indexed 0–9 from top to bottom, columns 0–9 from left to right. The border consists of all cells in row 0, row 9, column 0, and column 9. The grid uses colors 0–9 encoded as single characters; only G (green), B (blue), and . (black) appear in the input, and the output additionally uses R (red) and X (grey) in place of black cells.

Figurative Descriptions (16 variants)

"A walled city where open ground exposed to the wilderness beyond the walls catches fire, while sheltered courtyards remain in cool shadow."

120b: k=2

See: The grid is a fortified city seen from above. Green and blue tiles are the stone buildings and towers — permanent, immovable architecture. The dark empty spaces between them are open ground, bare earth where no structure stands. Some patches of open ground wind their way to the city walls at the grid's edge, forming paths that lead outside. Others are hidden courtyards, completely enclosed by buildings with no route to the perimeter.

Do: A wildfire sweeps in from the untamed lands beyond the city walls. Any open ground connected to the city's edge — where the flames can reach by traveling through adjacent clearings — ignites and burns red. But the courtyards sealed off deep within the city's stone heart are shielded from the inferno; they merely fill with ash-grey smoke, untouched by the flames themselves. The buildings stand fireproof, unchanged by the blaze.

Grid: The city is a 10×10 block. Its outer walls run along all four edges. Buildings appear as green and blue stones. The bare ground is black. Fire spreads only through orthogonal passages — it cannot leap diagonally between clearings. After the fire passes, exposed ground glows red and sheltered ground turns grey.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (10369 chars)

We need to infer rule: From description: green and blue are buildings (unchanged). Black '.' is open ground. Fire spreads from edge through orthogonal connectivity of '.' cells. Those reachable from edge become red 'R'. Those '.' cells not reachable (enclosed) become grey 'X' (ash). Buildings unchanged. Also note that in examples, some '.' become R, some become X. Also note that some '.' become R even if adjacent to edge via path through other '.' cells. Also note that some '.' become X even if adjacent to edge? Let's verify.

Example1 Input grid 10x10. Let's map '.' positions. Output shows R for many '.' reachable, X for enclosed. Also note that some '.' become R even if adjacent to edge? Edge cells: first row, first column, last row, last column. In example1, first row: ". G . G . . B G G B". Output first row: "R G R G R R B G G B". So '.' at (0,0) became R (edge). '.' at (0,2) became R (edge). '.' at (0,4) became R (edge). '.' at (0,5) became R (edge). So edge '.' become R.

Now interior '.' that are reachable via path become R. Enclosed become X. Let's test a known interior '.' that became X: In example1 row2 (index1) input: ". B B B B G . . B B". Output: "R B B B B G X X B B". So positions (1,0) edge '.' -> R. Positions (1,6) and (1,7) interior '.' became X. Are they enclosed? Let's see connectivity: They are adjacent to each other, but surrounded by B and G? Input row1 col6 '.' (index1,6). Adjacent orthogonal: up (0,6) is B (building), down (2,6) is G, left (1,5) is G, right (1,7) '.' . So no path to edge. So enclosed -> X.

Thus rule matches.

Now also note that some '.' become X even if adjacent to edge via other '.' that become X? No, they would become R if reachable.

Thus transformation: For each '.' cell, if there exists a path of '.' cells (including itself) to any edge cell (i.e., any '.' on border) via orthogonal moves, then change to R; else change to X. All other colors unchanged.

Now we need to apply to test input.

Let's parse test input 10x10.

I'll label coordinates (r,c) 0-9.

Row0: "B . . B . B B B B G"
Indices:
0 B
1 .
2 .
3 B
4 .
5 B
6 B
7 B
8 B
9 G

Row1: ". . . G . G . B . ."
0 .
1 .
2 .
3 G
4 .
5 G
6 .
7 B
8 .
9 .

Row2: ". B . G G . B G G G"
0 .
1 B
2 .
3 G
4 G
5 .
6 B
7 G
8 G
9 G

Row3: "G B G B B . G G . B"
0 G
1 B
2 G
3 B
4 B
5 .
6 G
7 G
8 .
9 B

Row4: "B B G . B G . B B ."
0 B
1 B
2 G
3 .
4 B
5 G
6 .
7 B
8 B
9 .

Row5: ". G . B G . B B . G"
0 .
1 G
2 .
3 B
4 G
5 .
6 B
7 B
8 .
9 G

Row6: "B B G . . G . G G G"
0 B
1 B
2 G
3 .
4 .
5 G
6 .
7 G
8 G
9 G

Row7: "G B B B B G B . G B"
0 G
1 B
2 B
3 B
4 B
5 G
6 B
7 .
8 G
9 B

Row8: "G . . . G G B . B B"
0 G
1 .
2 .
3 .
4 G
5 G
6 B
7 .
8 B
9 B

Row9: "B . G B B . . . B ."
0 B
1 .
2 G
3 B
4 B
5 .
6 .
7 .
8 B
9 .

Now we need to find '.' cells reachable from edge via '.' connectivity.

Edge cells are those on row0, row9, col0, col9.

Let's list all '.' on edges:

Row0: col1 '.' (0,1) edge. col2 '.' (0,2) edge. col4 '.' (0,4) edge. Others? col? col? (0,0) B, (0,3) B, (0,

... [7408 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 500: "A talent show lineup where the shyest performers shuffle to ... [10 chars truncated]" k=1 alts: 3/15 85b81ff1

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The grid is 13 rows by 14 columns. Columns 2, 5, 8, and 11 are fixed separator columns filled entirely with black (.). The remaining columns form five 2-column blocks: block 0 (cols 0-1), block 1 (cols 3-4), block 2 (cols 6-7), block 3 (cols 9-10), and block 4 (cols 12-13). Every even row (0, 2, 4, 6, 8, 10, 12) is fully filled with the grid's color in every block. Each odd row (1, 3, 5, 7, 9, 11) has each block in one of two states: 'full' (both cells are the color, e.g. M M) or 'dotted' (first cell is the color, second is black, e.g. M .). This creates a 6-row by 5-column binary matrix where 1 = full and 0 = dotted.

Do: 1. Identify the five 2-column blocks and the six odd rows. 2. For each block (column of the binary matrix), compute its column sum — the number of odd rows where that block is full (1). 3. Sort the five block-columns by their column sum in ascending order (fewest full blocks on the left, most on the right). Ties may be broken by preserving relative input order. 4. Reconstruct the grid: even rows remain fully filled across all blocks, separator columns remain black, and each odd row's block values come from the rearranged column order. The content within each block-column (which specific odd rows are full vs dotted) is preserved — only the left-to-right ordering of the block-columns changes.

Grid: The grid is always 13 rows by 14 columns. A single non-black color is used throughout (blue, orange, magenta, etc.). Columns 2, 5, 8, and 11 are always black separators. The grid has a rigid alternating structure: even rows are uniform, odd rows encode binary block states. The output grid has identical dimensions and structure.

Figurative Descriptions (16 variants)

"A talent show lineup where the shyest performers shuffle to the front."

120b: k=1

See: A stage is divided into five booths by black curtain dividers. Each booth has six spotlight moments (the odd rows) where performers can either stand tall and fill their booth with color, or shrink back into shadow, leaving a gap. Between these moments, the full chorus line blazes across every booth in unison. Each booth has its own personality — some performers are bold and fill nearly every spotlight, while others are timid, retreating into darkness more often than not.

Do: The stage director watches the dress rehearsal and counts how many times each booth's performer steps into the spotlight versus shrinks away. Then she rearranges the lineup: the shyest performer — the one who hid in shadow most often — is moved to the leftmost booth, and the boldest — the one who filled the most spotlights — takes the rightmost position. Each performer keeps their own unique pattern of courage and shyness; only their position on stage changes. The chorus-line moments remain untouched — everyone still shines together on those beats.

Grid: The stage is always thirteen rows deep and fourteen columns wide, lit in a single color against a black backdrop. Four black curtains hang at regular intervals, carving the stage into five equal booths. The chorus rows blaze uniformly; the solo rows reveal each booth's character — full presence or partial retreat. After the reshuffle, the stage dimensions and curtains remain identical, only the order of booths has changed.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (12330 chars)

We need to understand transformation. Input grid 13 rows x 14 columns. There are black curtains dividing into five booths. Black curtain likely columns of '.'? In examples, '.' is black. In input, pattern: columns: pattern repeats every? Let's examine example 1 input (not given but output shown). Actually we have only output for example 1. Need to infer rule: It says rearrange booths based on count of filled vs empty in odd rows (spotlight moments). The odd rows (1-indexed?) are rows 1,3,5,...? They are "solo rows" where each booth either full color or partial retreat. Even rows are chorus rows where all booths have uniform color across? In description: "Between these moments, the full chorus line blazes across every booth in unison." So even rows are uniform across all columns (same color). In example, let's check output rows: Row1: B B . B B . B B . B B . B B. That's pattern of booths: each booth maybe 2 columns? Let's see curtains: four black curtains dividing into five booths equal width. Total columns 14, so each booth width =? 14 columns, 4 curtains columns are black '.'? That would be 4 columns of '.' dividing, leaving 10 columns for booths, not equal. Maybe curtains are single column of '.' between booths. So 5 booths => 5*? +4 curtains =14 => booths width = (14-4)/5 =10/5=2 columns each. Yes each booth 2 columns, curtains are single column '.' between booths. Indeed pattern: B B . B B . B B . B B . B B. So each booth 2 columns of B (blue) in row1. Row2: B . . B . . B B . B . . B B. Let's parse: columns: 1 B,2 .,3 .,4 B,5 .,6 .,7 B,8 B,9 .,10 B,11 .,12 .,13 B,14 B. Curtain columns at positions 3,6,9,12? Indeed after each 2 columns, a '.' curtain column. Let's check: Booth1 columns1-2, curtain col3, booth2 col4-5, curtain col6, booth3 col7-8, curtain col9, booth4 col10-11, curtain col12, booth5 col13-14. Good.

Now rule: For each booth, count in odd rows (1,3,5,7,9,11,13) how many times the booth is "filled" vs "shrink". What defines filled? In odd rows, each booth either fully filled (both columns same color?) or partially empty (maybe one column '.'?). In example, look at odd rows: rows1,3,5,7,9,11,13. Row1: all booths fully B B. Row3 same as row1. Row5 same. Row7 same. Row9 same. Row11 same. Row13 same. So each booth always fully filled? Actually row2 is even (chorus) but not uniform? Row2 has pattern B . . B . . B B . B . . B B. That's not uniform across booths. Wait description says chorus rows blaze uniformly across every booth in unison. That would mean all booths same color across entire row? But row2 not uniform. Maybe I mis-index: maybe odd rows are the "chorus rows" (full across) and even rows are solo rows. Let's test: Row1 (odd) is B B . B B . ... not uniform across whole row because curtains are '.' but within booths both columns B. So across booths, same color B. So yes chorus rows have each booth filled with same color (maybe varying per test). Row2 (even) shows variation: some booths have B . (one filled one empty) e

... [9456 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 532: "A dancer performs before a triptych of mirrors, each catchin ... [29 chars truncated]" k=1 alts: 4/15 8e5a5113

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a 3×11 grid divided into three 3×3 panels separated by single columns of grey (X). The first panel contains a colored pattern using several distinct colors. The second and third panels are entirely black (empty), waiting to be filled.

Do: Copy the first 3×3 panel into panel 1 unchanged. Fill panel 2 by rotating the first panel 90 degrees clockwise: read the columns of panel 1 from bottom to top, left to right, so that column 0 (bottom to top) becomes row 0 of panel 2, column 1 (bottom to top) becomes row 1, and column 2 (bottom to top) becomes row 2. Fill panel 3 by rotating the first panel 180 degrees: reverse the order of the rows and within each row reverse the order of the cells, so the bottom-right cell of panel 1 becomes the top-left cell of panel 3.

Grid: The grid is 3 rows by 11 columns. It contains three 3×3 content panels at column positions 0–2, 4–6, and 8–10, separated by single-column grey (X) dividers at columns 3 and 7. The dividers are fixed and never change.

Figurative Descriptions (16 variants)

"A dancer performs before a triptych of mirrors, each catching her from a different angle."

120b: k=1

See: A dancer strikes a pose in the leftmost mirror of a triptych. The other two mirrors are dark, their silvered surfaces waiting to catch the light. Grey curtain-rods hang between the three frames.

Do: The second mirror doesn't face the dancer head-on — it watches from her side, catching the scene as if she'd spun a quarter-turn on her heel. What was at her feet now stretches across the top, read from left to right like a sentence. The third mirror hangs behind her: it sees everything the first mirror saw, but perfectly inverted — the dancer's reflection flipped upside-down and backwards, as though she gazed at herself from the opposite corner of the room.

Grid: The triptych hangs three square mirrors side by side, each three panes tall and three panes wide, separated by narrow grey curtain-rods that never move. Only the leftmost mirror holds the dancer's living colors; the other two must learn her shape by turning the light.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (3842 chars)

We need to infer rule. Input grid seems 3x? Actually examples: Input appears 3 rows, each row length? Example1 input: "M G G X . . . X . . ." that's 11 columns? Let's count: M G G X . . . X . . . = 11. So three mirrors side by side each 3x3 with grey curtain rods (X) separating. So pattern: 3 columns mirror1, then X column, then 3 columns mirror2, then X column, then 3 columns mirror3. So total 3*3 +2 =11 columns. Yes.

Thus each mirror is 3x3 region. Mirror1 contains original pattern (colors). Mirror2 and mirror3 initially empty (dots). Output shows transformed patterns placed into mirror2 and mirror3 according to description.

In example output: same dimensions 3 rows 11 columns. Mirror1 unchanged? Let's compare first row input mirror1: M G G ; output same M G G. Yes unchanged. Mirror2 (columns 5-7) output: first row "M M M". According to description: second mirror shows original rotated quarter-turn (90 degrees) clockwise? Let's test: Original mirror1 matrix:

Row1: M G G
Row2: M G G
Row3: M G R

Rotate 90 clockwise: becomes
Row1: M M M (from column1 bottom to top? Actually original column1: M,M,M becomes top row after rotation? Let's compute: original matrix A (r,c). Rotated 90 cw B(r,c)=A(3-c+1, r). So B row1 col1 = A(3,1)=M, row1 col2 = A(2,1)=M, row1 col3 = A(1,1)=M => M M M matches. Row2: B row2 col1 = A(3,2)=G, col2 = A(2,2)=G, col3 = A(1,2)=G => G G G matches output row2 mirror2: G G G. Row3: B row3 col1 = A(3,3)=R, col2 = A(2,3)=G, col3 = A(1,3)=G => R G G matches output row3 mirror2: R G G. Yes! So second mirror is 90-degree clockwise rotation of first mirror.

Third mirror: description says inverted (upside-down and backwards) i.e., rotated 180 degrees? Actually "flipped upside-down and backwards" suggests 180-degree rotation (or both vertical and horizontal flip). Let's see output mirror3 columns 9-11: first row: R G M. Original rotated 180 would be: original reversed both rows and columns: original matrix reversed rows and columns gives:

Original:
M G G
M G G
M G R

Rotate 180: row1 becomes original row3 reversed: R G M. Yes matches output row1 mirror3: R G M. Row2: original row2 reversed: G G M matches output row2 mirror3: G G M. Row3: original row1 reversed: G G M? Wait original row1 reversed: G G M. Output row3 mirror3: G G M. Yes matches. So third mirror is 180-degree rotation.

Thus rule: Input contains leftmost 3x3 pattern, other two are empty (dots). Output: copy original unchanged in leftmost; fill middle mirror with 90-degree clockwise rotation; fill rightmost mirror with 180-degree rotation.

Now test input:

Rows:
1: G G W X . . . X . . .
2: W W W X . . . X . . .
3: R W W X . . . X . . .

So leftmost 3x3 matrix:

Row1: G G W
Row2: W W W
Row3: R W W

We need to produce output with same dimensions 3x11.

Mirror1 unchanged.

Mirror2 (columns 5-7) = 90 cw rotation.

Compute rotation:

Original matrix A:

[ [G, G, W],
  [W, W, W],
  [R, W, W] ]

Rotate 90 cw: B rows:

Row1 = column1 bottom to top: A[3,1]=R, A[2,1]=W

... [869 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.03)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 535: "The kingdom with the most rebellions draws the chronicler's ... [10 chars truncated]" k=2 alts: 4/15 8efcae92

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black (.) background containing several distinct rectangular regions composed entirely of blue (B) and red (R) cells. Each region is a solid rectangle of B cells with some R cells scattered within it. The regions vary in size — some are small and some are large. They are separated from each other and from the grid edges by black cells.

Do: 1. Identify all distinct rectangular regions of non-black cells in the input grid. Each region is a contiguous rectangle of blue and red cells surrounded by black cells. 2. For each region, count the number of red (R) cells it contains. 3. Select the region that contains the greatest number of red cells. 4. Extract that region and output it as the result, trimmed to its exact bounding rectangle with no surrounding black cells.

Grid: The input is a 20×20 grid with a black (.) background. The output grid dimensions match exactly the height and width of the selected region. The output contains only blue (B) and red (R) cells — no black cells.

Figurative Descriptions (16 variants)

"The kingdom with the most rebellions draws the chronicler's attention."

120b: k=2

See: Across a dark sea, several blue kingdoms float as islands of varying sizes. Within each kingdom, red rebels hide among the loyal blue citizens — some kingdoms harbor just a handful of dissidents, while others are riddled with unrest.

Do: A royal chronicler surveys the archipelago from above, counting the rebels in each kingdom. The kingdom seething with the most red insurgents is the one worth documenting — it is lifted from the sea and preserved in the chronicle, its borders trimmed tight around its walls.

Grid: The dark sea fills a 20×20 world. The chronicled kingdom appears alone in the output, fitting exactly within its own fortress walls — only loyal blue citizens and red rebels remain, with no trace of the surrounding waters.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (11347 chars)

We need to infer rule: From description: dark sea = '.' background. Blue kingdoms = contiguous region of B? Within each kingdom, red R cells are inside. Need to find kingdom with most red cells, then output that kingdom trimmed to its bounding box, preserving B and R only (no '.'), maybe also keep shape exactly as in input (including interior '.'? but there are none inside kingdom? In examples, output contains only B and R, no '.'.

Let's analyze Example1. Input 20x20. Need to locate blue regions (connected orthogonally? likely 4-direction). There are many B clusters. The output is 9x9 grid. Let's find which region corresponds.

We can try to locate region with most R cells. Count R per B region.

But maybe simpler: The output appears to be a region extracted from input rows 2-10? Let's see output first row: B B B B R B B B B. In input, look at rows 2-? Actually row indices starting 0? Let's examine input snippet near top: rows 2-? Row2 (0-index) is ". . B B B B R B B B B . . . . . . . . .". That's row index 2. The segment columns? Let's write columns: positions 0-? The pattern: . . B B B B R B B B B . . . . . . . . . . So there is a B region from col2 to col10 inclusive (9 cells) with an R at col6? Actually col2 B,3 B,4 B,5 B,6 R,7 B,8 B,9 B,10 B. That's exactly B B B B R B B B B. That matches first row of output. Good.

Next output row: B R B B B B B B R. Look at input row3: ". . B R B B B B B B R . . . B B B B B R". The first region columns 2-10: B R B B B B B B R. matches output second row. So output corresponds to the region from rows 2-10, columns 2-10 (a 9x9 block) which is a blue kingdom with many reds.

Thus rule: Find the blue region (connected component of B) that contains the most R cells (i.e., count of R within that component). Then output the minimal bounding rectangle that contains that component, preserving B and R cells (and maybe '.' inside? but there were no '.' inside region). In example1, that region is 9x9.

Check example2. Input: there are two separate blue clusters? Let's locate region with most reds. Output is 5x6? Actually output grid lines: 5 rows, each 6 columns. Let's see if matches a region.

Find in input where there is a block of B with some Rs. Look at rows 13-17 (0-index). Row13: ". . . . . . . . . . . . . . . . . . . ." Actually row13 is all '.'? Wait example2 input lines: after many empty rows, there is a block starting at row13? Let's list rows with B:

Row13 (index 13?) Actually lines after many empty rows: row13 (0-index) maybe line 13: ". . . . . . . . . . . . . . . . . . . .". Not.

Let's examine given example2 input: after row 12 (0-index) there are many '.' rows. Then row13: ". . . . . . . . . . . . . . . . . . . .". Actually they show rows 13-? Let's parse:

Rows (starting 0):
0: . . . . . . . . . . . . . . . . . . . .
1: . . . . . . . . . . B B B B B B B B . .
2: . . . . . . . . . . B B B B R B B B . .
3: . . B B B B B . . . B B B B B B B B . .
4: . . B B B R B . . . B R B B B B R B . .
5: . . B B R B

... [8350 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.32)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 536: "A signet stamp presses its own seal wherever its ink touches ... [1 chars truncated]" k=2 alts: 0/15 8f2ea7aa

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a 9×9 grid with a black background. Somewhere within the grid, a small shape made of a single non-black color occupies a 3×3 region. This shape consists of some colored and some black cells within that 3×3 bounding box. The rest of the grid is entirely black.

Do: Divide the 9×9 grid into a 3×3 arrangement of 3×3 blocks (rows 0-2/3-5/6-8, cols 0-2/3-5/6-8). Extract the 3×3 shape from the block where it appears. Read the shape's binary pattern: each cell is either 'on' (colored) or 'off' (black). This binary pattern tells you which of the nine 3×3 blocks in the meta-grid should receive a copy of the shape. Specifically, if the shape's cell at relative position (r, c) is colored, then stamp a copy of the shape into the 3×3 block at meta-grid position (r, c). If the shape's cell is black, leave that block empty. The original block where the shape appeared is also governed by this rule — it keeps its copy only if the pattern dictates it.

Grid: The grid is always 9×9, naturally partitioned into a 3×3 meta-grid of 3×3 blocks. The background color is black (.). Only one non-black color is used. The output grid has the same 9×9 dimensions.

Figurative Descriptions (16 variants)

"A signet stamp presses its own seal wherever its ink touches."

120b: k=2

See: A 9×9 dark canvas is divided into a 3×3 mosaic of nine chambers. In one chamber, a small signet seal has been carved — a pattern of colored wax against dark stone, filling exactly one 3×3 chamber.

Do: The seal is both the message and the address book. Each dot of wax in the seal points to a chamber in the mosaic — the seal's own shape is a map of where it wants to be stamped. Wherever the wax touches within its tiny 3×3 frame, it dispatches a perfect replica of itself to the corresponding chamber of the larger mosaic. Where the seal is blank, that chamber stays dark and empty. The seal reproduces itself across the canvas according to its own image — it is a self-replicating blueprint.

Grid: The canvas is always 9×9, a mosaic of nine 3×3 chambers arranged 3×3. The night-dark background (.) is the unmarked stone. Only one color of wax is used. The final mosaic keeps the same 9×9 dimensions.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (9073 chars)

We need to infer rule. The description: 9x9 grid divided into 3x3 chambers each 3x3. There is a "seal" pattern of colored wax within one 3x3 chamber (i.e., a 3x3 subgrid containing the colored cells). That pattern is replicated to other chambers according to its own shape: each colored cell in the seal corresponds to a chamber location (i.e., which of the 9 chambers). For each colored cell in the seal, we place a copy of the entire seal pattern into the corresponding chamber in the larger grid. Where seal is blank, that chamber stays empty (i.e., all .). So we need to locate the 3x3 region that contains the colored cells (the "seal") in the input. Then replicate that pattern across the 9x9 grid: for each cell (i,j) within the 3x3 seal that is colored, we copy the entire 3x3 seal pattern into the chamber at position (i,j) in the 3x3 arrangement of chambers.

Let's test with Example 1.

Input example 1: we have O's in a small 3x3 region? Let's locate. Input grid shows O's at positions: row4 col6? Actually rows indexed 0-8. Input lines:

Row0: all .
Row1: all .
Row2: all .
Row3: ". . . . . O . . ." => O at (3,5)
Row4: ". . . . O O . . ." => O at (4,4) and (4,5)
Row5: ". . . O . . . . ." => O at (5,3)
others empty.

These O's form a shape within a 3x3 block? Let's see bounding box: rows 3-5, cols 3-5. That's a 3x3 block. Within that block, the pattern is:

Rows 3-5, cols 3-5:

Row3 (r=3): cols3-5 => positions (3,3) ., (3,4) ., (3,5) O => . . O
Row4: (4,3) ., (4,4) O, (4,5) O => . O O
Row5: (5,3) O, (5,4) ., (5,5) . => O . .

So seal pattern:
. . O
. O O
O . .

Now output example 1: we need to see if this pattern is replicated across chambers.

Output grid given: Let's map chambers. The output appears to have O's placed in a diagonal pattern across the whole grid. Let's compute expected: For each colored cell in seal (positions within 3x3), we copy the whole seal into the corresponding chamber.

First, label positions within seal (i,j) where i,j =0..2 (top-left of block). The colored cells are at (0,2), (1,1), (1,2), (2,0). So four cells.

Thus we will place copies of the seal into chambers at those positions in the 3x3 arrangement of chambers.

Chambers are arranged 3x3; each chamber is 3x3 cells. Chamber (0,0) corresponds to rows0-2, cols0-2; (0,1) rows0-2, cols3-5; etc.

So we need to place seal pattern into chambers (0,2), (1,1), (1,2), (2,0). Let's see output: O's appear in those chambers accordingly.

Check chamber (0,2): rows0-2, cols6-8. Output rows0-2, cols6-8 should contain the seal pattern. Output row0: "... ... O" at col8? Actually output row0: ". . . . . . . . O" => O at col8 (index8). Row1: ". . . . . . . O O" => O at col7,8. Row2: ". . . . . . O . ." => O at col6. This matches pattern . . O / . O O / O . . placed at (0,2) chamber (cols6-8). Indeed row0 col8 O (position (0,2) within chamber). Row1 col7 O (1,1) and col8 O (1,2). Row2 col6 O (2,0). Good.

Chamber (1,1): rows3-5, cols3-5. Output rows3-5, cols3-

... [6160 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 539: "A mountain casts its shadow from the corner where a sentinel ... [14 chars truncated]" k=2 alts: 1/15 902510d5

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains two kinds of colored elements against a black background. First, there is a single prominent shape made of one color, formed by the largest diagonally-connected component of any color. Second, there are several scattered 'marker' cells of various other colors (and possibly isolated cells of the shape's own color that are not connected to it). Exactly one marker sits on a corner of the grid. The remaining markers have varying colors, and one color among them is the most frequent.

Do: 1. Identify the main shape: find the largest connected component of any single color, using 8-connectivity (diagonal adjacency counts). Preserve this shape in the output exactly as it appears. 2. Identify all other non-black cells as markers. This includes cells of any other color AND any cells of the shape's color that are not part of its connected component. 3. Count the total number of markers: call this N. The triangle's leg length is N − 1. 4. Find the marker that sits exactly on a corner of the grid. This determines where the triangle is placed. Remove this marker. 5. Among the remaining markers, find the most frequent color. This is the triangle's fill color. Remove all markers from the grid. 6. Draw a filled right isosceles triangle with legs of length N − 1, right angle at the identified corner, extending inward along both grid edges. Specifically: at row/column distance d from the corner (d = 0 to N − 2), fill all cells from the corner along the perpendicular edge out to distance N − 2 − d. The triangle does not overwrite cells of the main shape.

Grid: Grids vary in size (not necessarily square), with black (.) as the background. The coordinate system is row-major: row 0 is the top, column 0 is the left. The four grid corners are (0,0), (0,W−1), (H−1,0), and (H−1,W−1). Exactly one marker always occupies one of these four corners.

Figurative Descriptions (16 variants)

"A mountain casts its shadow from the corner where a sentinel stands watch."

120b: k=2

See: Against the darkness, a great landmark stands — a constellation of connected cells in one color, the largest such formation on the field. Scattered around it are sentinel figures, lone colored cells standing at various posts. One sentinel has claimed a corner of the world as its watchtower. The others mill about the interior, and among them, one color appears more often than the rest — these are the shadow-bearers.

Do: The sentinel standing at the corner of the world sounds a horn and vanishes. All the other scattered figures dissolve as well, but not before their number is counted — their collective presence determines the size of the shadow they leave behind. The shadow takes the color of whichever sentinel tribe had the most members. From the now-empty corner, a great triangular shadow spreads inward along both walls, growing narrower as it reaches deeper into the grid — a right triangle whose legs stretch one fewer cell than the total number of sentinels that once stood. The ancient landmark is sacred ground; the shadow flows around it, never darkening its cells.

Grid: The world is a rectangular expanse of darkness. Its four corners are places of power — watchtowers where a sentinel may stand guard. Exactly one sentinel always occupies one of these watchtowers, and from that corner the shadow is cast inward, hugging both edges and tapering as it goes.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (13280 chars)

We need to decode puzzle. From description: "great landmark stands — a constellation of connected cells in one color, the largest such formation on the field." That's likely a connected component of same color (largest). In examples, we need to see.

Example1 input grid 6 rows? Let's list. Input rows:

Row1: . Y . . . . R . .
Row2: . . . . . R . . .
Row3: . . . . R . R . .
Row4: . . . . . . . R .
Row5: . . . Y . . . . .
Row6: . . . . . . . . G

Output rows:

Row1: . . . . . . R . .
Row2: . . . . . R . . .
Row3: . . . . R . R . .
Row4: . . . . . . . R .
Row5: . . . . . . . . Y
Row6: . . . . . . . Y Y

Observations: In input, there are Y at (1,2) and (5,4). R's many. G at (6,9). Output: Y's moved to bottom right corner area forming triangle? Let's see.

Sentinel corner: "Exactly one sentinel always occupies one of these watchtowers". Sentinel likely a single cell of a color that is not part of the landmark? In example1, which corner has a single colored cell? Input corners: top-left (1,1) is ., top-right (1,9) ., bottom-left (6,1) ., bottom-right (6,9) G. So G at bottom-right corner is sentinel. So sentinel at corner is G. It "sounds a horn and vanishes". So G disappears from corner. Indeed output has no G (except maybe other G? there were none else). So G removed.

Other scattered figures dissolve as well, but before they vanish, their number is counted — determines size of shadow they leave behind. The shadow takes the color of whichever sentinel tribe had the most members. So we need to count other colored cells that are "sentinels"? Actually "sentinel figures, lone colored cells standing at various posts." So these are isolated cells (not part of the landmark) maybe of any color. In example1, besides corner G, there are Y at (1,2) and (5,4). Are they isolated? Yes they are singletons not adjacent to same color. Also there is maybe other colors? Only R's are many and connected? R's form a shape maybe the landmark? Let's see connectivity: R cells positions: (1,7), (2,6), (3,5), (3,7), (4,8). Are they connected orthogonally? (1,7) adjacent to (2,7) none, (1,6) ., (1,8) ., (0,7) out. So not connected. Actually they are not orthogonally adjacent. Maybe connectivity includes diagonal? Not sure. But they might be considered separate. However description says "constellation of connected cells in one color, the largest such formation on the field." So we need to find largest connected component of any color. In example1, maybe the Y's are not connected, each isolated. R's also isolated. G only one. So largest component size is 1 (multiple). But maybe they consider connectivity including diagonal? Let's test: R at (1,7) and (2,6) are diagonal adjacent, maybe considered connected? If diagonal counts, then many R's could be connected forming shape. Let's see adjacency via 8-neighbor: (1,7) diagonal to (2,6) yes. (2,6) orthogonal to (3,6) none, diagonal to (3,5) yes. (3,5) orthogonal to (3,7) two apart, but diagonal to (4,6) none, but (3,7) diag

... [10319 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.02)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 543: "In a noisy room, the loudest voice shapes the echo on the wa ... [3 chars truncated]" k=2 alts: 15/15 9110e3c5

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Example 6

Input

→

Output

Example 7

Input

→

Output

Literal Description

See: The input is a 7×7 grid containing black (empty) cells and several different colors. One color appears far more frequently than all others — this is the dominant color. The remaining non-black colors (such as W, X, Y, O, A, M) each appear only once or twice, scattered throughout the grid as distractors.

Do: 1. Count the occurrences of each non-black color in the input grid. Identify the dominant color — the one that appears most frequently. 2. Map the dominant color to a fixed 3×3 output pattern using azure (A) on a black background: If the dominant color is R (red), output a horizontal line: row 1 = '. . .', row 2 = 'A A A', row 3 = '. . .'. If the dominant color is G (green), output an inverted-L shape: row 1 = '. A A', row 2 = '. A .', row 3 = '. A .'. If the dominant color is B (blue), output a mirrored shape: row 1 = '. . A', row 2 = 'A A .', row 3 = '. A .'. 3. Output the corresponding 3×3 pattern. The specific positions and arrangement of colors in the input grid do not matter — only which color is most frequent.

Grid: The input grid is always 7×7. The output grid is always 3×3. The output uses only two colors: black (.) for background and azure (A) for the pattern. The output pattern is entirely determined by which non-black color dominates the input grid, not by spatial arrangement.

Figurative Descriptions (16 variants)

"In a noisy room, the loudest voice shapes the echo on the wall."

120b: k=2

See: A crowded square buzzes with chatter. Most voices are faint — a whisper of maroon here, a murmur of orange there — but one color is practically shouting over everyone else, filling the room with its presence. The rest are just passersby.

Do: Listen to the crowd and find the voice that drowns out all others. That dominant voice doesn't just win the argument — it chooses a signature gesture, a pose struck in azure light on a small dark stage. Each color-voice has rehearsed its own distinctive pose: one stands bold as a horizon line, another rises like a bent elbow, another traces a zigzag path. The loudest voice steps onto the stage and strikes the only pose it knows.

Grid: The crowded square is always seven paces on each side. The stage where the winner performs is a tiny three-by-three platform, lit only in azure against darkness. No matter how chaotic the crowd, only one pose appears — chosen not by where anyone stood, but simply by who was loudest.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (3184 chars)

We need to infer rule. Input 7x7 grid, output 3x3 grid with only azure (A) and black (.) maybe? In examples output contains only '.' and 'A'. So rule: find dominant color (most frequent) in input, then output a specific pattern (pose) associated with that color, placed on 3x3 azure stage (i.e., azure cells are the pose? Actually output shows A where pose is, '.' elsewhere). So each color maps to a particular 3x3 pattern of A's (maybe shape). Need to deduce mapping from examples.

Example1 input: Let's count frequencies of each color (excluding .). Input grid 7x7.

I'll list rows:

1: B . B . O . .   => B,B,O
2: B B W B . B .   => B,B,W,B,B => B4,W1
3: . . B B . R .   => B,B,R
4: . . . . G . B   => G,B
5: . Y . B . . B   => Y,B,B
6: . . B . R . A   => B,R,A
7: . . B . O G B   => B,O,G,B

Now tally:
B: count? Row1 2, row2 4, row3 2, row4 1, row5 2, row6 1, row7 3 = total 15? Let's sum: 2+4=6, +2=8, +1=9, +2=11, +1=12, +3=15. Yes B=15.
O: row1 1, row7 1 =2.
W:1
R: row3 1, row6 1 =2.
G: row4 1, row7 1 =2.
Y:1
A:1

Dominant is B (blue). Output pattern corresponds to blue's pose.

Output grid (3x3):
. . A
A A .
. A .

Let's map coordinates (row,col) 0-index. Pattern of A's:
(0,2), (1,0),(1,1),(2,1). That's like a shape maybe like a bent elbow? Let's note.

Now Example2: Input grid.

Rows:
1: . G . G X G . => G,G,X,G => G3, X1
2: . . G G . . . => G,G => G2
3: A . . . . . G => A,G => A1,G1
4: G Y G W G . G => G,Y,G,W,G,G => G4,Y1,W1
5: . . W G B G G => W,G,B,G,G => G3,W1,B1
6: . G G G . G . => G4,G? actually row6: . G G G . G . => G4? there are G at col1,2,3,5 => 4 G.
7: . . . . . . G => G1

Now tally:
G: count rows: row1 3, row2 2 =>5, row3 1 =>6, row4 4 =>10, row5 3 =>13, row6 4 =>17, row7 1 =>18. So G=18.
X:1
A:1
Y:1
W:2? row4 W1, row5 W1 =>2
B:1

Dominant G (green). Output pattern:

. A A
. A .
. A .

A's at (0,1),(0,2),(1,1),(2,1). That's like a vertical line plus top right two? shape maybe "Z"? Not.

Now we need mapping for each possible dominant color to a pattern. We have only two examples: blue -> pattern1, green -> pattern2. Need to infer pattern for test input where dominant color maybe something else.

Let's compute frequencies for test input.

Test Input rows:

1: G . G . . . G => G,G,G => 3 G
2: G . W X . . X => G,W,X,X => G1,W1,X2
3: . G . G . R W => G,G,R,W => G2,R1,W1
4: A G . G . . O => A,G,G,O => A1,G2,O1
5: . G X . . G G => G,X,G,G => G3,X1
6: . . G G . . . => G,G => G2
7: . . G . Y . . => G,Y => G1,Y1

Now tally each color:

G: row1 3, row2 1 =>4, row3 2 =>6, row4 2 =>8, row5 3 =>11, row6 2 =>13, row7 1 =>14. So G=14.

W: row2 1, row3 1 =>2.
X: row2 2, row5 1 =>3.
R:1
A:1
O:1
Y:1

Dominant is G (green) again, same as example2. So output should be the pattern associated with green.

Thus output should be same as example2 output: pattern for green.

Pattern from example2:

Row0: 

... [343 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.78)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 561: "Two shields with holes reveal the vulnerable spots where arr ... [19 chars truncated]" k=1 alts: 2/15 94f9d214

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input is a rectangular grid with 8 rows and 4 columns, composed of two 4×4 halves stacked vertically. The top half contains a pattern made of green (G) cells and black (.) cells. The bottom half contains a pattern made of blue (B) cells and black (.) cells.

Do: Split the input grid into two 4×4 halves: the top half (green pattern) and the bottom half (blue pattern). For each cell position in the 4×4 grid, check whether that position is black (absent) in BOTH halves — that is, the cell is '.' in the top half AND '.' in the bottom half. If both are absent, place a red (R) cell at that position in the output. Otherwise, place a black (.) cell. The output is a single 4×4 grid.

Grid: The input grid is 4 columns wide and 8 rows tall, consisting of two equal 4×4 sub-grids stacked vertically. The output grid is 4×4. The background color is black (.).

Figurative Descriptions (16 variants)

"Two shields with holes reveal the vulnerable spots where arrows strike through."

120b: k=1

See: A warrior carries two shields stacked one behind the other — the front shield is forged from green iron, the rear from blue steel. Each shield has gaps where the metal was never filled in, leaving dark empty spaces.

Do: An archer fires arrows at the paired shields. Where both shields have holes — where neither green iron nor blue steel stands guard — the arrow passes clean through and leaves a red wound on the wall behind. Wherever at least one shield holds firm, the arrow is deflected and the wall stays unmarked.

Grid: The two shields sit one atop the other in the input, each a square slab four units across. The wall behind them, revealed in the output, is the same size as a single shield, dark except where it bleeds red.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (3969 chars)

We need to understand puzzle. Input consists of two 4x4 shields stacked vertically? In example, input grid appears 8 rows? Let's examine Example 1 input lines:

". G G ."
". G . G"
". . G ."
"G G G G"
"B B B B"
"B B . ."
"B B . ."
". B B ."

That's 8 rows, each 4 columns. So top 4 rows are green shield (G) with some '.' (black) holes. Bottom 4 rows are blue shield (B) with holes '.'.

Output is 4x4 grid:

". . . ."
". . R ."
". . . R"
". . . ."

So red appears where both shields have holes (i.e., positions where top shield has '.' and bottom shield has '.'), else '.' (black). Indeed check: Let's map coordinates (row0-3 top shield, row4-7 bottom). For each column.

Create matrix:

Top rows (0-3):
0: . G G .
1: . G . G
2: . . G .
3: G G G G

Bottom rows (4-7):
4: B B B B
5: B B . .
6: B B . .
7: . B B .

Now compute holes: top '.' positions: (0,0),(0,3),(1,0),(1,2),(2,0),(2,1),(2,3),(3 none). bottom '.' positions: (5,2),(5,3),(6,2),(6,3),(7,0). Also bottom row 7 col0 '.'.

Now intersection where both have '.' at same column row? Need align rows: The shields are stacked one atop the other, same size, so we compare each cell (i,j) where top row i (0-3) and bottom row i+4 (4-7). So pair rows: (0 with 4), (1 with5), (2 with6), (3 with7). Let's compute:

Row0 vs Row4: top row0: . G G . ; bottom row4: B B B B . No '.' in bottom row4, so intersection none -> output row0 all '.'.

Row1 vs Row5: top row1: . G . G ; bottom row5: B B . . . Intersection where both '.' at col2? top col2 '.' bottom col2 '.' yes. Also col3? top col3 G not '.'; col0 top '.' bottom B not '.'; col1 top G bottom B not '.' . So only col2 => output row1 col2 = R. Output row1: . . R . matches.

Row2 vs Row6: top row2: . . G . ; bottom row6: B B . . . Intersection at col2 '.' both? top col2 G not '.'; col3 top '.' bottom '.' yes at col3. Also col0 top '.' bottom B not '.'; col1 top '.' bottom B not '.' . So col3 => R. Output row2 col3 R matches.

Row3 vs Row7: top row3: G G G G ; bottom row7: . B B . . Intersection where top '.'? none, so all '.' output row3 all '.' matches.

Thus rule: output cell is red if both top and bottom corresponding cells are '.' (black). Else output '.' (black). So essentially logical AND of holes.

Now test input given:

Test Input:
". G . G"
"G G G ."
". . . G"
"G G G ."
". . B B"
". . B B"
". B . ."
"B B . ."

Again 8 rows 4 columns. Top shield rows0-3, bottom rows4-7.

Let's list rows:

Row0: . G . G
Row1: G G G .
Row2: . . . G
Row3: G G G .
Row4: . . B B
Row5: . . B B
Row6: . B . .
Row7: B B . .

Now compute output 4x4 where red where both top and bottom cells are '.'.

We need to pair rows: (0,4), (1,5), (2,6), (3,7).

Create columns 0-3.

Row0 vs Row4:
Row0: col0 '.'; col1 G; col2 '.'; col3 G
Row4: col0 '.'; col1 '.'; col2 B; col3 B

Intersection '.' both:
- col0: top '.' bottom '.' => red.
- col1: top G not '.' => no.
- col2: top '.' bottom B not '.' => no.
- col3: top G not '.' => no.

So output row0: R at col0, others '.'.

Row1 v

... [1029 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 584: "A river finds its way down a mountainside, flowing straight ... [70 chars truncated]" k=1 alts: 7/15 992798f6

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid has a black background with exactly two colored pixels: one red (R) and one blue (B), placed at different positions. No other non-black pixels exist.

Do: Draw a green path connecting R and B (the endpoints themselves remain their original colors; only the intermediate cells become green). To construct the path: (1) Compute the vertical distance dy and horizontal distance dx between R and B. (2) The path is built starting from R toward B. Take exactly one diagonal step (moving one cell toward B both vertically and horizontally). (3) Next, draw a straight segment in whichever axis has the greater distance (vertical if dy > dx, horizontal if dx > dy). This straight segment has length |dy - dx| and is placed immediately after that first diagonal step (i.e., near the R end of the path). (4) Finally, continue with diagonal steps (each moving one cell toward B in both row and column) until reaching B. The total path uses min(dy, dx) diagonal steps and |dy - dx| straight steps. The straight portion always sits near R's end of the path.

Grid: Rectangular grid of arbitrary dimensions (observed up to 16x16). Background is black (.). The grid uses a standard row-column coordinate system with (0,0) at top-left. The grid dimensions and positions of R and B vary across instances. The output grid has the same dimensions as the input, with the only change being the addition of green (G) cells forming the connecting path.

Figurative Descriptions (16 variants)

"A river finds its way down a mountainside, flowing straight where the slope is steep and cutting diagonally across gentle terrain."

120b: k=1

See: Two signal fires burn on a dark plain — one a red flame and one a blue flame — separated by empty night. They are beacons waiting to be linked.

Do: A green vine grows from the red flame toward the blue, seeking connection. The vine prefers to travel diagonally — cutting across the field at a slant — but where the distance in one direction far exceeds the other, it must first run straight like a river in a canyon, hugging close to the red flame's side before it can angle freely. The vine always gets its straight running out of the way early, near the red end, then sweeps diagonally the rest of the journey until it reaches the blue flame. The flames themselves are never overgrown — they remain their original colors, proud endpoints of the green bridge between them.

Grid: A dark rectangular plain of arbitrary size. Two lone colored embers — one red, one blue — sit somewhere on the field. After the vine grows, every cell it passes through turns green, but nothing else on the plain changes.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (13473 chars)

We need to infer rule. From description: two points red and blue. Draw a green path from red to blue. Preference: diagonal, but if distance in one direction far exceeds other, must first run straight near red, then diagonal rest. Essentially path is like: start at red, move straight (horizontal or vertical) for some steps to reduce the larger delta, then move diagonally (i.e., one step both row and column) until reaching blue. The straight segment is near red end. The diagonal segment continues to blue.

In example 1: Input grid 15x15. Red at (row1? Let's index from 0). Let's map coordinates: rows 0-14, columns 0-14.

Red at row1 col6 (since line 2 has R at column 6). Blue at row12 col3 (line 13 has B at col3). So red (1,6), blue (12,3). Compute delta rows = 12-1 = 11 downwards. delta cols = 3-6 = -3 leftwards. Absolute values: dr=11, dc=3. Larger is dr. According to rule, first run straight (vertical) near red to reduce dr until dr equals dc? Actually they said "where the distance in one direction far exceeds the other, it must first run straight like a river in a canyon, hugging close to the red flame's side before it can angle freely." So we need to run straight for the excess difference: excess = dr - dc = 8? Wait dr=11, dc=3, difference =8. So we should move vertically 8 steps first, then diagonal for remaining 3 steps (since after reducing dr to 3, both equal). Let's test output: In output, green cells from (2,5?) Let's list.

Output shows green column at col5? Actually output rows 2-11 have G at column5? Let's examine: Row2 (index2) has G at column5? Input row2 (index2) originally all . . . . . . . . . . . . . . . . . Output row2: ". . . . . G . . . . . . . . . .". That's column5 (0-index). Row3 same column5. Row4 same column5. Row5 same column5. Row6 same column5. Row7 same column5. Row8 same column5. Row9 same column5. Row10 same column5. Row11 has G at column4? Wait row11 output: ". . . . G . . . . . . . . . . .". That's column4. So path: Starting at red (1,6). Then green at (2,5) (down-left diagonal). Actually row2 col5 is diagonal step (down-left). Then continues down-left diagonal until row10 col5? Wait row10 col5 is same column5, row10 index10. Let's compute: Starting (1,6). Diagonal down-left moves to (2,5), (3,4), (4,3), (5,2), (6,1), (7,0) would go out of bounds. But output shows vertical line at column5 not diagonal. Let's examine more carefully.

Maybe we misread coordinates. Let's list output rows with indices:

Row0: all .
Row1: R at col6.
Row2: G at col5.
Row3: G at col5.
Row4: G at col5.
Row5: G at col5.
Row6: G at col5.
Row7: G at col5.
Row8: G at col5.
Row9: G at col5.
Row10: G at col5.
Row11: G at col4.
Row12: B at col3.

So path: from red (1,6) go straight down (vertical) at column5? Wait red column6, but green column5. So first step maybe left (horizontal) then vertical? Actually there is a shift left by 1 at start, then vertical down column5, then at row11 shift left again to column4 then down to blue at (12,3) which 

... [10509 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.64)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 596: "A reflection pool mirrors what stands at its edge, rippling ... [20 chars truncated]" k=2 alts: 6/15 9b30e358

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input grid is divided into two regions: an upper region filled entirely with a single background color, and a lower region containing a multi-colored pattern anchored to the bottom of the grid. The pattern region consists of the bottommost rows that contain at least one non-background color. The background color is the color that fills the upper region completely.

Do: 1. Identify the background color: the color that fills the top rows uniformly. 2. Identify the pattern: the contiguous block of rows at the bottom of the grid that contain non-background pixels. Call this block P, with P[0] being the topmost pattern row and P[n-1] being the bottommost. 3. Count the number of blank rows above the pattern (rows filled entirely with the background color). 4. Fill the blank rows by tiling the pattern upward, cycling through the pattern rows in reverse. Starting from the blank row immediately above the pattern, assign it the last row of P, the next blank row up gets the second-to-last row of P, and so on. When you exhaust the pattern rows, wrap around and continue from the last row of P again. 5. Keep the original pattern rows in place at the bottom.

Grid: The grid dimensions vary across instances. The width and height are not fixed. The output grid has the same dimensions as the input grid. The pattern occupies the bottom portion and the remaining rows above are filled by tiling the pattern upward with wraparound.

Figurative Descriptions (16 variants)

"A reflection pool mirrors what stands at its edge, rippling upward into the sky."

120b: k=2

See: A quiet, empty expanse of still water stretches across the upper portion of the world. At the shore along the bottom edge, a colorful monument stands — a small sculptural arrangement of vivid tiles rising from the earth. The water above is perfectly calm, showing nothing yet, waiting to be disturbed.

Do: The monument at the water's edge gazes down into the pool and sees its own reflection climbing upward through the stillness. The reflection doesn't simply mirror once — it keeps echoing, like a stone's ripple that never fades. The monument's image scrolls upward through the water, its bottom reaching up first, then its middle, then its top, and when the whole shape has appeared, it begins again from the bottom, endlessly tiling its echo toward the horizon. The original monument never moves — it remains planted at the shore while its ghost fills every empty inch of sky-water above.

Grid: The world varies in size but the shore is always at the bottom. The still water is one uniform color. The monument is the cluster of living colors rooted to the bottom edge. After the reflection, every row of water has been claimed by the monument's upward echo, and the world is the same size as before.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (12302 chars)

We need to infer rule. From description: bottom edge contains monument (colored cluster). Water is uniform color (single color) above. The output fills water rows with repeated upward scrolling of monument pattern.

Look at Example1. Input grid 10 rows? Let's count: lines: X X X X X (row1) ... up to row10? Actually there are 10 rows: rows 1-5 all X, row6: X R R R X, row7: X X R X X, row8: X A A X X, row9: X X A A X, row10: X X A X X. So bottom rows (6-10) contain colored pattern (R and A) forming a shape. The water color is X (grey). The monument is the cluster of non-X colors at bottom edge (rows 6-10). The output: rows 1-5 now contain the pattern repeated upward, shifting? Output rows: first row is X R R R X (which was row6 originally). second row X X R X X (row7). third row X A A X X (row8). fourth row X X A A X (row9). fifth row X X A X X (row10). Then rows 6-10 repeat same pattern again (same as rows1-5). So essentially the pattern of bottom 5 rows is tiled vertically to fill entire grid, preserving order. So the water region (originally rows1-5) is replaced by repeating the monument pattern upward, starting from bottom pattern's first row (closest to water) placed at topmost water row, then continue upward, then after reaching top, continue again? Wait we have 5 water rows, pattern height =5 (rows6-10). They tiled exactly once. If water rows > pattern height, they'd repeat multiple times. Example2: Input grid 10 rows, 7 columns. Water color is G (green). Bottom rows (7-10) contain pattern with W and R colors. Let's list rows 1-6 all G. Row7: G G G W R W G
Row8: G G G R W R G
Row9: G G G W W W G
Row10: G G G G W G G

So pattern height =4 rows (rows7-10). Output: rows1-10 show pattern tiled? Let's see output rows:

Row1: G G G W W W G
Row2: G G G G W G G
Row3: G G G W R W G
Row4: G G G R W R G
Row5: G G G W W W G
Row6: G G G G W G G
Row7: G G G W R W G
Row8: G G G R W R G
Row9: G G G W W W G
Row10: G G G G W G G

We can see pattern of 4 rows appears repeated twice (rows1-4 then rows5-8 then rows9-10? Actually pattern of 4 rows repeated twice gives 8 rows, but we have 10 rows. Let's examine pattern sequence: original pattern rows7-10 are:

A: W R W
B: R W R
C: W W W
D: G W G (but note row10 has G G G G W G G? Wait row10 input: G G G G W G G. That's pattern row D: G W G with extra G at start? Actually pattern includes the surrounding G columns (since monument is only colored tiles not G). In input, the whole row is G G G G W G G (so only one W tile at column5). So pattern rows include the background G as water? Wait water color is G, but monument includes colored tiles (W,R) on top of G background? In example1, water color X, but monument rows also have X background with colored R/A. So the pattern includes the background color as part of rows. So the pattern is the entire rows (including water color) of bottom region.

Thus pattern rows are exactly rows7-10 as given. Output rows1-4 correspond to rows9-10? Let's map:

Output row1: G G G W W 

... [9326 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 598: "Magnets repel — red pieces flee from the border like o ... [28 chars truncated]" k=1 alts: 12/15 9b4c17c4

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The grid is divided into two rectangular regions by background color: one region filled with A (azure) and one region filled with B (blue). The two regions share a boundary and together tile the entire grid (they may be side-by-side or stacked). Scattered within each region are one or more R (red) shapes — contiguous rectangular blocks of red cells.

Do: For each region, determine which edge of the region is the boundary shared with the other region, and which edge is the outer (non-shared) edge. Slide every R shape within that region horizontally or vertically — away from the shared boundary and toward the outer edge of its region — until the shape touches the outer edge. Each R shape preserves its size, orientation, and row/column position along the axis perpendicular to the slide direction. Specifically: if the two regions are stacked vertically, R shapes slide horizontally — in the A region they slide to the left edge, in the B region they slide to the right edge. If the regions are side by side, R shapes slide horizontally toward the far edge of their own region (away from the boundary between A and B). The background cells vacated by each shape return to the region's background color.

Grid: The output grid has the same dimensions as the input grid. The two background regions (A and B) remain in exactly the same positions. Only the R (red) cells move; no cells change color except by R cells vacating old positions (restored to background) and occupying new positions.

Figurative Descriptions (16 variants)

"Magnets repel — red pieces flee from the border like opposite poles pushing apart."

120b: k=1

See: The grid is a world split into two kingdoms — one azure, one blue — sharing a tense border. Scattered across each kingdom are red refugees, huddled in small rectangular groups, caught between the contested frontier and the distant outer walls of their homeland.

Do: The border between the kingdoms is electrified, radiating a repulsive force. Every red group feels the push and flees from the shared border, sliding straight toward the farthest wall of its own kingdom until it presses flat against the outer edge. They keep their shape and their lane — they don't drift sideways, only retreat as far from the charged border as their kingdom's walls allow.

Grid: The two kingdoms hold their ground — azure stays azure, blue stays blue. Only the red refugees move, and the ground they abandon heals back to the local kingdom's color. The world's dimensions never change.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (11677 chars)

We need to understand the puzzle. The description: two kingdoms azure (A) and blue (B) share a border. Red groups are rectangular groups (clusters) that are between the border and outer walls. They slide away from the border toward the farthest wall of its own kingdom, keeping shape, moving straight (presumably orthogonal direction) until they hit outer edge. The ground they leave becomes the kingdom's color (A or B). So basically, for each contiguous rectangular block of R that lies within a region of A or B, we need to move it away from the border between A and B, i.e., towards the outer edge of that region.

We need to identify the border: the line separating A and B. In the grid, A and B are separate contiguous regions? In example, the top part is A rows, bottom part B rows. The border is a horizontal line between them. Indeed example input: first 8 rows are A, then rows 9-12 are B. So border is between row 8 and 9 (horizontal). Red groups appear in both kingdoms: some in A region (rows 3-4, columns 3-4 etc) and some in B region (rows 6-7, columns 6-8 etc). They slide away from border: For A region (above border), they move upward (toward top edge). For B region (below border), they move downward (toward bottom edge). They keep shape and column positions? Actually they slide straight (vertical) while staying in same columns (lane). So they shift vertically until they hit the outer edge (top row for A, bottom row for B). The output shows that.

Let's verify with example: Input A region rows 1-8 (0-index?). Red block at rows3-4 (0-index maybe 2-3) columns3-4 (2-3). In output, those Rs moved to rows1-2? Actually output shows rows3-4 become A, rows1-2 become R? Let's examine output lines:

Output rows (1-index):
1: A A A A A A A A A A A
2: A A A A A A A A A A A
3: R R A A A A A A A A A
4: R R A A A A A A A A A
5: A A A A A A A A A A A
6: R R R A A A A A A A A
7: R R R A A A A A A A A
8: A A A A A A A A A A A
9: B B B B B B B B B B B
10: B B B B B B B B B R R
11: B B B B B B B B B R R
12: B B B B B B B B B B R

Let's map original red blocks:

Block1: rows3-4 cols3-4 (2x2). In output they moved to rows3-4? Wait they are now at rows3-4 cols1-2 (shifted left?). Actually output shows R at rows3-4 cols1-2. That's not vertical shift; it's horizontal shift leftwards. Hmm maybe border is vertical? Let's examine original grid more: Input rows:

Rows 1-8 are A region. Row1: A A A A A A A A A A A
Row2: same
Row3: A A R R A A A A A A A
Row4: A A R R A A A A A A A
Row5: A A A A A A A A A A A
Row6: A A A A A R R R A A A
Row7: A A A A A R R R A A A
Row8: A A A A A A A A A A A
Rows9-12 are B region.

So there are two red blocks in A region: a 2x2 at (row3-4, col3-4) and a 2x3 at (row6-7, col6-8). In output, the first block moved to rows3-4 col1-2 (shift left). The second block moved to rows6-7 col1-3 (shift left). So they moved leftwards, not up. So border must be vertical between A and B? But A region is left side? Wait A region is top part, B bottom. However th

... [8704 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 602: "A paint roller loaded with a pattern stamps its design acros ... [36 chars truncated]" k=1 alts: 12/15 9c1e755f

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains one or two L-shaped structures made of grey (X) cells against a black (.) background. Each L-shape consists of two perpendicular arms meeting at a corner. One arm is a straight line of grey cells (the 'length arm'). The other arm (the 'pattern arm') sits at the end of the length arm and contains either: (a) a single column or row of colored cells adjacent to the grey arm, where each cell is a different color — this is a 1D color sequence; or (b) a small rectangular block of colored cells (2 or more rows/columns) forming a 2D tile pattern. The L-shape defines a rectangular fill region: the area swept out by extending the pattern arm along the full length of the length arm.

Do: For each L-shaped grey structure in the grid: 1. Identify the two arms of the L. One arm is a straight bar of grey cells. The other arm meets it at a corner. 2. Determine which arm carries the pattern. The pattern arm has colored (non-black, non-grey) cells adjacent to it, either as a 1D sequence along the arm or as a 2D block at the end of the arm. The other arm (the length arm) defines how far the pattern should be replicated. 3. Count the number of cells along the length arm — this is how many times (or how far) the pattern must be extended. 4. If the pattern is 1D (one colored cell per row or column along the pattern arm): extend each color across the full span defined by the length arm. Each colored cell becomes a full row or column of that color, matching the length of the length arm. 5. If the pattern is 2D (a small rectangular tile of colored cells): tile that pattern repeatedly along the direction of the length arm to fill the rectangular region. The tile repeats cyclically until the full length is covered. 6. The grey cells and all other cells outside the fill region remain unchanged.

Grid: The grid is 10x10 with a black (.) background. Grey (X) cells form L-shaped borders that remain in place in the output. The fill regions are rectangular areas bounded by the L-shapes. Cells outside the L-shapes and their fill regions stay black. Multiple independent L-shapes may exist in the same grid, each processed separately.

Figurative Descriptions (16 variants)

"A paint roller loaded with a pattern stamps its design across a wall measured by its guide rail."

120b: k=1

See: On a dark canvas, grey steel scaffolding forms right-angled brackets — like an architect's L-square laid down. One arm of each bracket is a measuring rail, counting off how far the work must go. Where the other arm meets it, a painter has loaded a swatch — sometimes a single stripe of colors stacked like paint samples on a card, sometimes a small patchwork tile like a fabric sample. The swatch waits at the corner, ready to be rolled out across the empty wall the bracket frames.

Do: The painter picks up the swatch from the corner of the bracket and loads it onto a roller. Then, guided by the measuring rail, the roller presses the swatch across the blank wall — one stamp for every rung on the rail. If the swatch is a single column of colors, each color bleeds sideways into a full band stretching the length of the rail. If the swatch is a patchwork tile, it prints and reprints in a wallpaper pattern, tiling end to end until the rail runs out. The scaffolding itself is structural — it never takes paint, never moves. It just frames the workspace and dictates the reach of the roller.

Grid: The dark canvas is 10×10. Grey scaffolding brackets stand rigid, forming the borders of rectangular workspaces. The painter only fills inside the workspace each bracket defines — everything beyond the scaffolding stays bare. If two brackets appear on the same canvas, each painter works independently within their own frame.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (15324 chars)

We need to understand the puzzle. The description: grey scaffolding brackets form right-angled brackets (L shape). One arm is a measuring rail (counting off how far the work must go). The other arm meets it at corner where a swatch (single column of colors stacked, or a patchwork tile) is placed. Then the painter picks up the swatch and rolls it across the empty wall inside the bracket, one stamp per rung on the rail. If swatch is a single column of colors, each color bleeds sideways into a full band stretching the length of the rail. If swatch is a patchwork tile, it prints and reprints in a wallpaper pattern, tiling end to end until the rail runs out.

So we need to detect brackets formed by X (grey). The bracket is an L shape: vertical arm and horizontal arm meeting at a corner. The vertical arm is the measuring rail? Or the horizontal? The description: "One arm of each bracket is a measuring rail, counting off how far the work must go. Where the other arm meets it, a painter has loaded a swatch". So the other arm (the one not the rail) holds the swatch at the corner. The rail determines length of fill.

In example 1: Input grid shows X's forming a horizontal line from (1,1?) Let's index rows 0-9, columns 0-9.

Example 1 Input:

Row0: . . . . . . . . . .
Row1: . X X X X X X . . .
Row2: . R . . . . . . . .
Row3: . B . . . . . . . .
Row4: . R . . . . . . . .
Row5: . M . . . . . . . .
Row6: . B . . . . . . . .
Row7: . B . . . . . . . .
Row8: . . . . . . . . . .
Row9: . . . . . . . . . .

So X's are at row1 columns1-6 (six X's). That's a horizontal line. No vertical line? Actually there is no vertical X's. So bracket? Maybe the vertical arm is the column of X's? But we only have horizontal. Wait maybe the bracket is formed by X's that are contiguous forming a right angle: a horizontal line and a vertical line meeting at a corner. In example 1, we have only horizontal line. But maybe the vertical arm is implied by the column of X's at column1? Let's check column1 rows1-? Row1 col1 is X, row2 col1 is R (not X). So no vertical X.

Thus maybe the bracket is just a line of X's that serves as rail, and the swatch is at one end (the leftmost X?) Actually the swatch appears as colored cells adjacent to the X line? In example, we have colored cells (R,B,R,M,B,B) placed in column2? Let's see coordinates: Row2 col1 is R? Actually row2 col1 is R (since row2: . R . . . . . . . .). That's column1 (0-index). Row3 col1 B, row4 col1 R, row5 col1 M, row6 col1 B, row7 col1 B. So there is a vertical column of colors directly below the X line? The X line is at row1 columns1-6. The colored column is at column1 rows2-7. So the corner is at (row1,col1) where X meets the column of colors? The swatch is a vertical column of colors (R,B,R,M,B,B). That's a single column swatch. The rail is the horizontal line of X's extending to the right from the corner. So the rule: replicate the column of colors horizontally across the length of the rail (excluding the corner X maybe). Ou

... [12330 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.02)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 611: "A stained-glass window reflected in a still pool of water" k=1 alts: 3/15 9f236235

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a large grid divided into a smaller grid of rectangular cells by single-width separator lines of a uniform color (e.g., azure or red). Each cell is a square block of identical pixels — either filled with a non-separator, non-black color (like blue, green, red) or empty (all black/dot). The separator lines form a complete lattice of horizontal and vertical bands, creating an N×M arrangement of equally-sized cells.

Do: 1. Identify the separator color — the color that forms complete horizontal and vertical lines spanning the full grid width/height. 2. Use these separator lines to segment the grid into a matrix of rectangular cells. 3. Reduce each cell to a single value: if the cell is filled with a non-separator color, record that color; if the cell is all black (dots), record black (dot). This produces a small abstract grid. 4. Flip this abstract grid horizontally — reverse the order of columns (mirror left to right). 5. Output the resulting flipped grid.

Grid: The output grid has dimensions equal to the number of cell-rows by the number of cell-columns in the input's lattice structure. Each pixel in the output represents one cell from the input. The background color is black (dot). The output contains only the cell-fill colors and black — the separator color does not appear in the output.

Figurative Descriptions (16 variants)

"A stained-glass window reflected in a still pool of water"

120b: k=1

See: A cathedral window is divided into panes by leading — thin, unbroken veins of a single bold color that run from edge to edge, both across and down, forming a lattice. Each pane between the leading is either filled with a jewel-tone color or left as clear dark glass.

Do: Stand before the window and read each pane as a single gemstone — sapphire, emerald, ruby, or empty night. Arrange these gems into a mosaic matching the window's layout. Now walk to the reflecting pool on the opposite side of the chapel: the water mirrors the window, swapping left for right, so what hung on the eastern edge now shimmers on the west. Record what the water shows you.

Grid: The reflection is a small mosaic — one stone per pane. The leading itself dissolves in the water and never appears in the reflection. Only the jewel colors and the darkness between them remain, painted in the pool's quiet reversal.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (9035 chars)

We need to infer rule from example. Input is large grid with leading lines? The description: leading are thin veins of single bold color that run across and down, forming a lattice. They are removed in output, only colors inside panes remain, reflected horizontally (mirror left-right). So we need to identify the "leading" color: likely azure 'A' maybe? In example, input shows many A forming grid lines (vertical and horizontal). The output is 4x4 grid (since there are 4 panes?). Let's examine example 1.

We have input grid size? Let's count rows: appears 19 rows (since pattern). Output is 4 rows, 4 columns.

The leading likely are the A's forming a grid of 5x5? Let's think: The input seems to have a pattern: top left area of '.' (black) then A column? Actually first rows: ". . . . A R R R R A . . . . A . . . ." repeated 4 rows. So there is a vertical line of A at column 5, then a block of R's, then A at column 10, then block of '.' then A at column 15 etc. So A's are vertical lines. Also there are horizontal lines of A rows: row 5 is all A's. Row 10 is ". . . . A . . . . A B B B B A . . . ." etc. Row 15 is all A's. Row 20? Actually there are 19 rows, with A rows at 5,10,15? Let's see: after row 4, row5 is all A's. Then rows 6-9 have pattern with A at start and after B block etc. Row10 is ". . . . A . . . . A B B B B A . . . ." etc. Row11-14 similar. Row15 is all A's. Row16-19 have pattern with G etc. So the A's form a grid dividing into 4x4 panes (since there are 3 vertical lines and 3 horizontal lines, making 4x4 cells). The colors inside each pane are either a solid color (R, B, G) or '.' (black). The output is 4x4 where each cell corresponds to a pane, showing the dominant color? Let's map.

Pane positions: top-left pane (rows 1-4, cols 1-4) are '.' (black). Output top-left is '.'? Output first row first column is '.' (they show .). Indeed output first row: ". . R ." So first cell '.' matches.

Top row second pane (rows1-4, cols6-9) contains R's. Output first row second column is '.'? Wait output first row second column is '.' (they have .). Actually output first row: ". . R ." So columns: 1 '.'; 2 '.'; 3 'R'; 4 '.'. So the pane with R's appears at column3. Let's map: The vertical A lines at columns 5,10,15. So panes are columns 1-4 (pane1), 6-9 (pane2), 11-14 (pane3), 16-19 (pane4). The R block is at columns 6-9 (pane2). So output should have R at column2, but they have at column3. Hmm maybe they also reflect horizontally (mirror). So after reflection, column order reversed: pane1 <-> pane4, pane2 <-> pane3. So R originally in pane2 becomes pane3 after reflection. Indeed output column3 is R. Good.

Now top row third pane (cols11-14) is '.' (since there is A at col10 then '.' area). Output column2 is '.' (makes sense after reflection). Top row fourth pane (cols16-19) is '.'? Actually there is '.' after A at col15? In first rows, after A at col15 there are '.' . So pane4 is '.'; after reflection becomes column1 '.' (output column1 '.

... [6059 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.25)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 648: "A still lake reflects the sky, but where the reflection meet ... [34 chars truncated]" k=3 alts: 8/15 a8610ef7

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains cells of two types: azure (A) cells and black (.) cells. The azure cells form an irregular pattern scattered across the grid, while black cells serve as the background.

Do: For each cell in the grid: (1) If the cell is black (.), leave it unchanged. (2) If the cell is azure (A) at position (row, col), find its vertical mirror position at (H-1-row, col), where H is the grid height. Check what occupies that mirror position in the input grid. If the mirror position also contains azure (A), replace the current cell with red (R). If the mirror position contains black (.), replace the current cell with grey (X).

Grid: The grid is 6 rows by 6 columns. Black (.) is the background color and remains unchanged throughout the transformation. The coordinate system has row 0 at the top. The vertical reflection axis is the horizontal midline of the grid, pairing row r with row H-1-r.

Figurative Descriptions (16 variants)

"A still lake reflects the sky, but where the reflection meets solid ground, it turns to stone."

120b: k=3

See: A landscape of azure figures stands on a dark stage. Some areas are empty darkness, others are occupied by azure presences — scattered unevenly, as if a crowd milling about before a performance.

Do: A great lake lies along the horizon, perfectly still. Each azure figure gazes down into the water to find its reflection on the opposite shore. If the figure sees a twin gazing back — another azure soul standing at the mirrored spot — both blush red with recognition. But if the figure peers into the water and finds only empty darkness where a twin should be, it turns grey with loneliness, hardening to stone. The dark empty spaces, having no one to cast a reflection, remain untouched.

Grid: The stage is six steps wide and six steps tall. The lake's surface runs horizontally through the middle, so each figure on the upper shore has a counterpart position on the lower shore, and vice versa. The darkness is merely empty ground — it plays no part in the reflection and stays as it is.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (5969 chars)

We need to infer rule. From description: azure figures = A. Dark empty = . (black). Lake is horizontal middle line: each figure on upper shore has counterpart position on lower shore (mirrored vertically across middle). If both have A at mirrored positions, they both become red (R). If only one side has A and opposite side is ., then that A becomes grey (X). The dark empty spaces remain unchanged (stay .). Also note that after transformation, the lower side also changes accordingly: if upper A has counterpart A, both become R. If upper A has no counterpart, it becomes X, and lower side . stays .? Wait description: "If the figure sees a twin gazing back — another azure soul standing at the mirrored spot — both blush red with recognition. But if the figure peers into the water and finds only empty darkness where a twin should be, it turns grey with loneliness, hardening to stone. The dark empty spaces, having no one to cast a reflection, remain untouched." So for a solitary A, it becomes X. The counterpart is . (dark) stays . (unchanged). For a pair of As, both become R.

Thus rule: For each cell (i,j) where i is row index (0-5). Mirror row = 5 - i (since 6 rows). For each pair (i,j) and (mirror_i,j). If both are A -> both become R. Else if one is A and other is . -> the A becomes X, . stays . . If both . -> stay . . Also note there may be As on both sides but after processing one side, the other also processed; but we must ensure not double change. So final output: For each column j, look at the two rows (top and bottom). Determine.

Let's test with Example1 to verify.

Example1 Input grid 6x6:

Row0: . A . A A A
Row1: A A A A A .
Row2: A . A . A .
Row3: . A A A . A
Row4: A A . A A .
Row5: A A . . . A

Mirror pairs: row0 with row5, row1 with row4, row2 with row3.

Compute column-wise.

I'll compute row0 vs row5:

col0: row0 '.' , row5 'A' => solitary A at row5 becomes X? Output row5 col0 is X (check output row5: X R . . . R) yes col0 X. row0 stays '.' (output row0 col0 '.'), matches.

col1: row0 'A', row5 'A' => both become R. Output row0 col1 R, row5 col1 R. Yes.

col2: row0 '.' , row5 '.' => both '.' output row0 col2 '.' row5 col2 '.'? Output row5 col2 '.' yes.

col3: row0 'A', row5 '.' => solitary A at row0 becomes X, row5 stays '.' . Output row0 col3 X, row5 col3 '.'? Output row5 col3 '.' yes.

col4: row0 'A', row5 '.' => solitary A at row0 X, row5 '.' . Output row0 col4 X, row5 col4 '.' yes.

col5: row0 'A', row5 'A' => both R. Output row0 col5 R, row5 col5 R. Yes.

Now row1 vs row4:

Row1: A A A A A .
Row4: A A . A A .

col0: A & A => both R. Output row1 col0 R, row4 col0 R. Yes.

col1: A & A => both R. Output row1 col1 R, row4 col1 R. Yes.

col2: A & . => solitary A at row1 becomes X, row4 stays . . Output row1 col2 X, row4 col2 '.'? Output row4 col2 '.' yes.

col3: A & A => both R. Output row1 col3 R, row4 col3 R. Yes.

col4: A & A => both R. Output row1 col4 R, row4

... [3158 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.31)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 652: "Azure is a disease — too much infection and the host b ... [17 chars truncated]" k=1 alts: 0/15 a934301b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a black (.) background containing several distinct rectangular objects made of blue (B) cells. Embedded within some of these blue rectangles are one or more azure (A) cells, which replace blue cells at certain positions. Each object is a contiguous rectangular blob of blue and azure cells, separated from other objects by at least one column or row of black cells.

Do: 1. Identify each distinct rectangular object composed of blue (B) and azure (A) cells. Objects are separated by black (.) background cells. 2. For each object, count the number of azure (A) cells it contains. 3. If an object contains 0 or 1 azure cells, keep it in place unchanged. 4. If an object contains 2 or more azure cells, remove it entirely by replacing all of its cells with black (.) background. 5. All kept objects remain at their original positions; the rest of the grid stays black.

Grid: The output grid has the same dimensions as the input grid. The background is black (.). The grid coordinate system is row-major with (0,0) at the top-left. Objects do not overlap and do not touch the edges of other objects.

Figurative Descriptions (16 variants)

"Azure is a disease — too much infection and the host body is condemned."

120b: k=1

See: Against a dark void, blue organisms float like cells under a microscope. Some cells appear perfectly healthy — pure blue tissue. Others show signs of azure infection, where the pathogen has replaced healthy blue tissue with azure lesions.

Do: A quarantine protocol sweeps the grid. Each organism is examined for infection. Those with only a single azure lesion — or none at all — have a strong enough immune response to survive and remain in place. But organisms riddled with two or more azure lesions are too far gone; the quarantine team erases them completely, leaving only darkness where they once lived.

Grid: The petri dish remains the same size after the purge. Surviving organisms hold their original positions, untouched and unaware of their neighbors' fate. The void claims everything else.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (11677 chars)

We need to infer rule. From description: blue organisms (B) are cells, azure infection (A). Rule: For each organism (connected component?) those with only a single azure lesion — or none — survive; those with two or more azure lesions are erased (set to .). Output shows many rows cleared after certain rows.

In example 1, input grid 14x14? Output shows only top 5 rows unchanged, rest all '.'.

Let's analyze example 1: Input shows many B and A scattered. Output retains same top 5 rows exactly same as input. Rows 6-14 become all '.'.

Thus rule likely: Identify each "organism" maybe a contiguous region of B (blue) cells? The azure lesions are A cells within that region. If a region has <=1 A, keep region unchanged; else erase entire region (turn all its cells to .). In example, rows 6-14 contain B and A but perhaps they belong to regions that have >=2 A, thus removed. Let's test: Row6: ". . . . . . . . . B B A B ." contains B B A B. That region maybe connected to other B's above? Let's see connectivity: likely orthogonal adjacency (4-direction). The top region of B's in left side maybe separate from right side region. Row6's B's at columns 10-12 maybe part of region that extends upward? Look at rows 1-5: there are B's in columns 10-13 etc. Row1: "... B B ." etc. So there is a region of B's in right side that continues down. That region includes many A's? Let's count A's in that region across whole grid: In rows 1-5 there are some A's: row1 col6? Actually row1: ". . . . . B A B . . . B B ." includes A at col7? Row1 col7 is A. Row2: ". . . . . B B B . . . B B ." no A. Row3: "B B B B . B B B . . . B B ." no A. Row4: "B A B B . . . . . . . . . ." includes A at col2. Row5: "B B B B . . . . . . . . . ." no A. So region maybe includes those A's. Count A's in that region: row1 col7, row4 col2. That's 2 A's. So region has >=2 A, thus should be erased. Indeed rows 1-5 in output still have those B and A? Wait output rows 1-5 unchanged, still have A's. So my hypothesis wrong.

Maybe the "organism" is each individual B cell considered as organism, and azure lesions are adjacent A cells? But description: "Each organism is examined for infection. Those with only a single azure lesion — or none at all — have a strong enough immune response to survive and remain in place. But organisms riddled with two or more azure lesions are too far gone; the quarantine team erases them completely". So each organism maybe a connected component of B cells (blue). Azure lesions are A cells that are adjacent (maybe orthogonal) to that component? If component has >=2 A cells adjacent (touching) then component removed.

In example, top region (rows1-5) maybe has only 0 or 1 adjacent A? Let's examine adjacency: The A cells are within the region (i.e., part of component?) Actually A cells are azure lesions inside the organism, maybe they replace B cells. So they are part of same component? If component includes both B and A cells (blue tissue and azu

... [8743 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.02)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 655: "A beam of light finds the longest path through a fractured s ... [10 chars truncated]" k=4 alts: 4/15 aa300dc3

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: A 10x10 grid bordered entirely by grey (X) cells. Inside the border, there is an irregular cavity of black (.) cells surrounded by grey (X) cells that protrude inward from the border, creating a jagged opening. The cavity is connected and roughly spans from one corner region of the interior to the opposite corner region.

Do: Draw a diagonal line of azure (A) cells through the open cavity. Consider both diagonal directions: top-left to bottom-right (\) and top-right to bottom-left (/). For each direction, find the starting cell and trace the diagonal one step at a time (moving +1 row and +1 or -1 column per step). A diagonal begins at the first black (.) cell encountered along the appropriate edge of the cavity and continues as long as each successive cell is black (.). The diagonal stops just before it would land on a grey (X) cell. Compare the two candidate diagonals and choose the one that is longer. Replace each black cell along the chosen diagonal with azure (A). All other cells remain unchanged.

Grid: Grid is always 10x10. Grey (X) cells form the complete outer border (all of row 0, row 9, column 0, column 9) plus irregular inward protrusions that shape the cavity. The interior contains only black (.) cells before transformation. After transformation, some black cells become azure (A) along exactly one diagonal line. Coordinates are (row, column) with (0,0) at top-left.

Figurative Descriptions (16 variants)

"A beam of light finds the longest path through a fractured stone wall."

120b: k=4

See: A fortress of grey stone encloses a dark cavern. The walls are uneven—chunks of stone jut inward like broken teeth, leaving a jagged hollow of shadow (black) within. The cavern stretches loosely from one corner of the chamber toward the opposite.

Do: A beam of azure light seeks passage through the darkness. It can only travel diagonally, slicing corner to corner like sunlight through a cracked ceiling. The light tests both slants—descending left and descending right—but stone snuffs it out on contact. Whichever angle lets the beam travel farther before striking wall, that is the path it takes. Where the light touches shadow, it leaves a glowing azure trace. The stone and the remaining darkness are unmoved.

Grid: The fortress is always a 10×10 chamber. Grey stone forms every outer wall and reaches inward in jagged formations, sculpting the shape of the cavern. The cavern itself is pure darkness (black) before the light arrives. After the beam passes, a single diagonal streak of azure marks its journey through the dark. Nothing else changes.

gpt-oss-120b — MARC valid, solved with 4 examples

Input

→

Prediction

vs

Expected

Reasoning (11741 chars)

We need to infer rule. From description: grey stone X forms outer wall and jagged interior, black . is cavern. Azure A is beam path diagonal from some start point on edge? Actually beam originates from "azure light seeks passage through the darkness. It can only travel diagonally, slicing corner to corner like sunlight through a cracked ceiling. The light tests both slants—descending left and descending right—but stone snuffs it out on contact. Whichever angle lets the beam travel farther before striking wall, that is the path it takes. Where the light touches shadow, it leaves a glowing azure trace. The stone and the remaining darkness are unmoved."

Interpretation: Starting point likely at a corner of the cavern? Possibly at a point where black region touches outer wall? The beam starts from a point on the boundary of black region (cavern) that is adjacent to stone? Actually "beam of azure light seeks passage through the darkness. It can only travel diagonally, slicing corner to corner". It tests both slants—descending left and descending right—but stone snuffs it out on contact. Whichever angle lets the beam travel farther before striking wall, that is the path it takes. Where the light touches shadow, it leaves a glowing azure trace."

Thus we need to find a starting black cell that is adjacent to stone? In examples, azure A appears as a diagonal line from some point near outer wall into interior, ending before hitting X.

Let's examine Example1 input and output to see pattern.

I'll map coordinates (row, col) 0-index.

Example1 Input grid: 10x10.

I'll write rows:

0: X X X X X X X X X X
1: X . . . . . X X X X
2: X X . . . . X . . X
3: X . . . . . . . . X
4: X X . . . . . . . X
5: X X . . . . . . X X
6: X . . . . . . . . X
7: X . . X X . . . . X
8: X X X X X . X X . X
9: X X X X X X X X X X

Output same but with A's:

Row1: X A . . . . X X X X (A at (1,1))
Row2: X X A . . . X . . X (A at (2,2))
Row3: X . . A . . . . . X (A at (3,3))
Row4: X X . . A . . . . X (A at (4,4))
Row5: X X . . . A . . X X (A at (5,5))
Row6: X . . . . . A . . X (A at (6,6))
Row7: X . . X X . . A . X (A at (7,7))
Row8: X X X X X . X X A X (A at (8,8))

So diagonal from (1,1) to (8,8). Starting point (1,1) is a black cell originally? Input row1 col1 was '.' yes black. It is adjacent to outer wall? At (1,1) its up neighbor (0,1) is X (stone). left neighbor (1,0) is X. So it's a black cell that is adjacent to stone on two sides (corner). The beam travels diagonally down-right (descending right) until just before hitting stone at (9,9) which is X (outer wall). Actually (8,8) is black originally? Input row8 col8 is '.'? Row8: "X X X X X . X X . X" So col8 (0-index) is '.' yes black. Next diagonal step would be (9,9) which is X, so stops before stone, leaving A at (8,8). So path is maximal diagonal from a starting black cell that is adjacent to stone (touching at corner?) The beam chooses direction (down-right or down-left) whichever yields longer pat

... [8798 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 4 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 659: "A bartender pours drinks into five reserved glasses, each gl ... [74 chars truncated]" k=1 alts: 1/15 aaecdb9a

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a 10x10 grid with an orange (O) background containing scattered blobs of up to five non-orange colors: grey (X), red (R), azure (A), maroon (W), and magenta (M). Each color appears as one or more distinct connected components (clusters of same-colored cells connected horizontally, vertically, or diagonally via 8-connectivity). Some components are compact shapes (crosses, rectangles, L-shapes), while others are single isolated cells or irregular groups.

Do: 1. Identify every non-orange color present in the input. The five possible colors each have a fixed column assignment in the output: grey (X) → column 0, red (R) → column 1, azure (A) → column 2, maroon (W) → column 3, magenta (M) → column 4. 2. For each color, count the number of distinct 8-connected components (groups of same-colored cells where each cell touches at least one other cell in the group horizontally, vertically, or diagonally). 3. Let the maximum component count across all five colors be H. The output grid has H rows and 5 columns. 4. Fill the output with orange (O), then for each color, fill its assigned column from the bottom row upward for as many rows as that color's component count. Colors not present in the input get height 0 (their column stays all orange).

Grid: The input is always a 10x10 grid with orange (O) as the background color. The output is an H×5 grid, where H equals the largest number of 8-connected components found for any single color. The output background is orange (O). Columns are indexed 0–4 from left to right, each permanently assigned to one color: col 0 = grey, col 1 = red, col 2 = azure, col 3 = maroon, col 4 = magenta.

Figurative Descriptions (16 variants)

"A bartender pours drinks into five reserved glasses, each glass filling according to how many parties of that color showed up tonight."

120b: k=1

See: An orange dance floor hosts gatherings of up to five different colored crowds—grey, red, azure, maroon, and magenta. Some crowds cluster into tight formations like crosses or blocks, while others splinter into loners scattered across the floor. Each color may have thrown one big party or fragmented into several smaller cliques that don't touch each other, even at the corners.

Do: The bartender has five glasses lined up on the counter, each permanently reserved for one color. At closing time, she counts how many separate cliques each color brought to the dance floor—friends who stuck together count as one group, even if they're only touching diagonally. She then pours into each glass one unit for every clique that color fielded. The tallest glass determines how tall the counter display is; shorter pours just leave empty space above. Colors that never showed up get nothing in their glass.

Grid: The dance floor is always a 10×10 venue awash in orange. The counter display is a lineup of exactly five glasses arranged left to right—grey's glass, red's glass, azure's glass, maroon's glass, magenta's glass—each filled from the bottom up. The display is only as tall as the fullest glass, and any unfilled space above a pour remains orange, like air above the liquid line.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (11696 chars)

We need to understand puzzle. Input grid 10x10 orange background with colored cells (M,R,A,W etc). Output is a small grid of 5 columns (glasses) and height equal to max number of connected components per color? Let's examine example.

Example 1 Input grid 10x10 with colors: W (maroon), M (magenta), R (red). Output grid shown:

Output:
O O O O M
O R O W M

That's 2 rows, 5 columns. Columns correspond to order: grey (X), red (R), azure (A), maroon (W), magenta (M). In example, there is no grey or azure, so those columns maybe empty? Output shows first column O (orange) for both rows, meaning no fill in grey glass. Second column has R in bottom row only (since red had 1 component?). Third column O both rows (no azure). Fourth column W in bottom row only (maroon). Fifth column M in both rows (magenta). So height is 2 rows because max components among colors is 2 (magenta had 2 components). Let's verify.

Count components in input example 1.

Let's map positions. Input grid lines:

Row0: O O O O O O O O W W
Row1: O O M O O O O O W W
Row2: O M M M O O O W O O
Row3: O O M O O O W O O O
Rows4-5 all O.
Row6: R O R O O O O M O O
Row7: O R O O O O M M M O
Row8: R O R O O O O M O O
Row9: O O O O O O O O O O

We need to count connected components for each color, using connectivity? The description: "friends who stuck together count as one group, even if they're only touching diagonally." So connectivity includes diagonal adjacency (8-direction). So components are 8-connected.

Now count components:

Red (R): positions: (6,0), (6,2), (7,1), (8,0), (8,2). Let's see connectivity: Are they all connected via 8-neighbor? Let's map coordinates (row,col). (6,0) adjacent to (7,1) diagonally, also (6,2) is two columns away not adjacent. (7,1) adjacent to (8,0) diagonal and (8,2) diagonal? (8,2) is two rows down and two columns right from (6,0) not directly. But via (7,1) they connect. Also (6,2) adjacent to (7,1) diagonal, and to (8,2) vertical? (8,2) is two rows down, same column, distance 2 rows, not adjacent. But (6,2) adjacent to (7,1) diagonal, so all reds are connected via (7,1). So single component. So red count =1.

Maroon (W): positions: (0,8),(0,9),(1,8),(1,9),(2,7),(3,6). Let's see connectivity: (0,8) adjacent to (0,9) horizontally, also (1,8) diagonal, etc. (2,7) is at row2 col7, adjacent to (1,8) diagonal, so connected. (3,6) adjacent to (2,7) diagonal. So all W are connected as one component. So maroon count=1.

Magenta (M): positions: (1,2), (2,1),(2,2),(2,3),(3,2),(6,7),(7,6),(7,7),(7,8),(8,7). Let's see connectivity: The cluster at top left (rows1-3) are connected together. The cluster at bottom right (rows6-8) are connected together. Are they connected via diagonal? Distance between (3,2) and (6,7) far, not. So two separate components. So magenta count=2.

Thus max components =2, matches output height 2 rows.

Now output grid: each column corresponds to color order: grey (X), red (R), azure (A), maroon (W), magenta (M). For each column, we fill from b

... [8732 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.03)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 660: "A pebble dropped in a still pond sends out rings of color." k=1 alts: 12/15 aaef0977

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input is an 8×8 grid filled entirely with orange (O) cells except for exactly one cell that has a different color — the 'seed.' The seed can be any of the nine non-orange colors: black (.), blue (B), red (R), green (G), yellow (Y), grey (X), magenta (M), orange is excluded as it is the background, azure (A), or maroon (W).

Do: 1. Locate the single non-orange cell in the input. Record its color C and its position (r, c) using zero-indexed row and column coordinates. 2. Use the fixed 9-color cycle in this order: G, Y, ., X, R, A, W, M, B (indices 0 through 8). 3. Determine the index of the seed color C within this cycle. Call it i. 4. For every cell (row, col) in the output grid, compute the Manhattan distance d = |row − r| + |col − c|. 5. Assign that cell the color at position (i + d) mod 9 in the cycle. 6. Fill the entire grid this way. The result is a diamond-shaped ripple of colors radiating outward from the seed position.

Grid: The grid is always 8×8. The input uses orange (O) as the uniform background with exactly one non-orange seed cell. The output grid has the same 8×8 dimensions but contains no orange — every cell is replaced by a color from the 9-color cycle determined by its Manhattan distance from the seed.

Figurative Descriptions (16 variants)

"A pebble dropped in a still pond sends out rings of color."

120b: k=1

See: A calm, uniform orange lake stretches across the grid — perfectly still, without a single ripple. Somewhere in this lake, a single colored pebble rests just beneath the surface, waiting to disturb the peace.

Do: The pebble drops. From the point of impact, concentric diamond-shaped waves radiate outward, each ring shifting to the next color in an endless chromatic heartbeat — a prayer wheel of nine hues spinning forever forward. The closer a spot is to where the stone fell, the younger its ripple; the farther away, the more the wave has aged and cycled through the spectrum. The particular color of the pebble determines which note the melody begins on, and every ring thereafter follows the same ancient song in order.

Grid: The once-tranquil orange lake is completely transformed — no trace of stillness remains. Every point on the water now shimmers with the color of whichever ripple has reached it, painting a diamond-shaped mandala centered on the point of impact. The lake keeps its original shores but is reborn entirely in the nine sacred colors of the wave.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6513 chars)

We need to infer rule. Input grid: mostly O (orange) with a single '.' (black) at position (row3,col2) using 1-index? Let's map example.

Example1 Input: 8x8 grid. All O except a G at (1,7?) Let's see: Input lines:
Row1: O O O O O O G O (col7 G)
Rows2-8 all O.

Output 8x8 with many colors. Looks like a diamond pattern centered at the G? The description: lake orange, pebble colored determines start note. In example, pebble is G (green). So center is G. Then concentric diamond rings with colors cycling through nine hues in order. Need order of colors: given color key includes 9 colors? list: . black, B blue, R red, G green, Y yellow, X grey, M magenta, O orange, A azure, W maroon. That's 10 actually. But description says nine hues. Maybe black is not part of cycle? In output we see colors: W, A, R, X, ., Y, G, etc. So cycle includes W, A, R, X, ., Y, G, B, M? Let's see order from center outward.

Find center at (1,7) where input G. Output at same position? Output row1 col7 is G? Row1: W A R X . Y G Y => col7 = G indeed. So center retains G (starting color). Then ring distance 1 (Manhattan distance) should be next color in cycle. Let's compute positions with Manhattan distance 1 from (1,7): (1,6), (1,8), (0,7 invalid), (2,7). Output values: (1,6)=Y, (1,8)=Y, (2,7)=? Row2: M W A R X . Y . => col7 = Y. So distance1 color = Y. So after G comes Y.

Distance2 positions: Manhattan distance 2 diamond. Let's list some: (1,5), (1,9 invalid), (2,6), (2,8), (3,7). Output values: (1,5)=. ; (2,6)=. ; (2,8)=. ; (3,7)=? Row3: B M W A R X . X => col7 = . . So distance2 color = . (black). So after Y comes . .

Distance3: positions distance3: (1,4), (2,5), (3,6), (4,7), (3,8), (2,9 invalid), (5,7) etc. Let's check some: (1,4)=X, (2,5)=X, (3,6)=X, (4,7)=? Row4: G B M W A R X R => col7 = X. So distance3 color = X (grey). So after . comes X.

Distance4: color? positions distance4: (1,3)=R, (2,4)=R, (3,5)=R, (4,6)=R, (5,7)=? Row5: Y G B M W A R A => col7 = R. So distance4 color = R (red). So after X comes R.

Distance5: color? (1,2)=A, (2,3)=A, (3,4)=A, (4,5)=A, (5,6)=A, (6,7)=? Row6: . Y G B M W A W => col7 = A. So distance5 color = A (azure). After R comes A.

Distance6: color? (1,1)=W, (2,2)=W, (3,3)=W, (4,4)=W, (5,5)=W, (6,6)=W, (7,7)=? Row7: X . Y G B M W M => col7 = W. So distance6 color = W (maroon). After A comes W.

Distance7: color? (2,1)=M, (3,2)=M, (4,3)=M, (5,4)=M, (6,5)=M, (7,6)=M, (8,7)=? Row8: R X . Y G B M B => col7 = M. So distance7 color = M (magenta). After W comes M.

Distance8: color? (3,1)=B, (4,2)=B, (5,3)=B, (6,4)=B, (7,5)=B, (8,6)=B, (9,7 invalid). Row8 col6 = B. So distance8 color = B (blue). After M comes B.

Distance9: color? (4,1)=G, (5,2)=G, (6,3)=G, (7,4)=G, (8,5)=G. Row8 col5 = G. So distance9 color = G again? That would complete cycle of 9 colors: starting G, then Y, ., X, R, A, W, M, B, then back to G. That's 9 steps? Let's count: G (0), Y (1), . (2), X (3), R (4), A (5), W (6), M (7), B (8) then repeats G (9). T

... [3855 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 674: "Loyal pets hear their owner's whistle and come running home." k=1 alts: 5/15 ae3edfdc

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains exactly two 'center' pixels — one blue (B) and one red (R) — placed on a black background. Scattered around the grid are 'satellite' pixels: orange (O) pixels that share a row or column with the blue center, and green (G) pixels that share a row or column with the red center. Each satellite is aligned with its center in one of the four cardinal directions (directly above, below, left, or right), separated by some distance of black cells. There may also be stray colored pixels that do not align cardinally with either center; these are noise.

Do: 1. Identify the two center pixels: blue (B) and red (R). Note their positions. 2. For each center, find its associated satellites — orange (O) for blue, green (G) for red — that lie in the same row or same column as that center (i.e., directly above, below, left, or right of it at any distance). 3. For each satellite, determine its cardinal direction relative to its center (up, down, left, or right). 4. Clear the entire grid to black. 5. Place each center pixel back at its original position. 6. Place each satellite in the cell immediately adjacent to its center in the same cardinal direction it originally occupied. That is, if a satellite was anywhere above its center, place it one cell above; if it was to the left, place it one cell to the left; and so on.

Grid: The grid is 15 rows by 15 columns. The background is black (.). The output grid has the same dimensions. Only the center pixels and their collapsed satellites appear in the output; all other cells are black.

Figurative Descriptions (16 variants)

"Loyal pets hear their owner's whistle and come running home."

120b: k=1

See: Two homeowners live on a vast dark field — one in a blue house, one in a red house. Each owner has pets of a signature color: the blue owner keeps orange cats, and the red owner keeps green dogs. The pets have wandered off along the roads (rows and columns) leading away from home, some near, some far. A few stray animals roam the field too, but they belong to nobody.

Do: Each owner stands on the porch and calls their pets home. Every pet hears the whistle and bolts straight back along the road it wandered down, not stopping until it reaches the yard — curling up in the very first patch of grass next to the front door, on whichever side it came running from. The strays, hearing no familiar voice, simply vanish into the night. When morning comes, each home stands with its faithful companions huddled as close as possible, one on each side they returned from.

Grid: The dark field is 15 plots by 15 plots. Two homes — blue and red — anchor the scene. After the calling, only the homes and their returned pets remain; the rest of the field is empty darkness.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6461 chars)

We need to infer rule from example. Input grid 15x15. There are two homes: blue house B and red house R. Pets: orange cats O belong to blue, green dogs G belong to red. They are placed somewhere along rows/columns emanating from the house? In example, we see B at (row3 col4?) Let's index rows starting 0. Input example1 grid: I'll map coordinates.

I'll rewrite input example1 rows with indices.

Row0: ". . . O . . . . . . . . . . ."
Row1: all .
Row2: "O . . B . . . . . O . . . . ."
Row3: all .
Row4: all .
Row5: all .
Row6: all .
Row7: all .
Row8: ". . . O . . . . . . . . . . ."
Row9: all .
Row10: ". . . . . . G . . . . R . . ."
Row11: all .
Row12: all .
Row13: all .
Row14: ". . . . . . . . . . . G . . ."

Output example1:

Row0: all .
Row1: ". . . O . . . . . . . . . . ."
Row2: ". . O B O . . . . . . . . . ."
Row3: ". . . O . . . . . . . . . . ."
Rows4-9 all .
Row10: ". . . . . . . . . . . G R . ."
Row11: ". . . . . . . . . . . G . . ."
Rows12-14 all .

Observations: In input, B at (2,3). O's at (0,3), (2,0), (2,9), (8,3). Also O at (6,0) and (6,14) in row6? Wait row6: "O . . . . B . . . . . . . . O" yes O at (6,0) and (6,14). That's far from B. In output, O's have moved to positions adjacent to B along same row or column direction they came from. Let's see: B at (2,3). Output O's at (2,2) left side, (2,4) right side, (1,3) above, (3,3) below. That's exactly one step away in each cardinal direction where there was at least one O somewhere along that line in input. Indeed there were O's in same column above (0,3) and below (8,3) and also maybe below further? Also O's in same row left (2,0) and right (2,9). So they returned to immediate neighbor cells.

Similarly for R at (10,11). Input G's at (10,6) same row left, (14,11) same column below, maybe also (0,11?) No. Also G at (0,11?) Actually row0 col11 is '.'; row0 col? there is G at row0 col11? Row0: ". . . O . . . . . . . . . . ." No G. Row14 col10? row14 has G at col10? Row14: ". . . . . . . . . . . G . . ." That's col11? Let's count: indices 0-14. Row14: positions: 0 .,1 .,2 .,3 .,4 .,5 .,6 .,7 .,8 .,9 .,10 .,11 G,12 .,13 .,14 . So G at (14,11) same column below R. Also maybe G at (10,6) left. Output shows G at (10,10) left neighbor? Wait output row10: ". . . . . . . . . . . G R . ." So G at col10 (one left of R at col11). Row11: ". . . . . . . . . . . G . . ." That's G at (11,11) below R (one down). So matches.

Other stray O's at (6,0) and (6,14) are not aligned with B (row6 not same row/col). So they vanish. Similarly stray G's not aligned vanish.

Thus rule: For each house (B or R), look along its row and column for same-colored pet (O for B, G for R). If any pet of that color exists in that direction (any distance), then place a pet of that color adjacent to the house on that side (up/down/left/right) in output. Remove all other elements; only houses and possibly adjacent pets remain. If multiple pets in same direction, still only one adjacent pet.

Also note that if there are multiple house

... [3470 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.01)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 696: "Helium balloons rising by the breath it took to fill them" k=1 alts: 2/15 b2bc3ffd

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid has a fixed row of azure (A) cells along the bottom. Above this azure row, one or more distinct colored shapes (connected groups of non-orange, non-azure cells) rest on or near the azure floor. The rest of the grid is filled with orange (O) background cells.

Do: For each colored shape in the grid, count its total number of cells (its area). Then move that entire shape upward by exactly that many rows. The azure bottom row remains fixed in place. All vacated cells become orange background. Shapes maintain their exact internal arrangement — only their vertical position changes.

Grid: The grid is 8 columns wide and 8 rows tall. The bottom row is always entirely azure (A). The background color is orange (O). Shapes occupy the rows above the azure floor and are composed of non-orange, non-azure colored cells.

Figurative Descriptions (16 variants)

"Helium balloons rising by the breath it took to fill them"

120b: k=1

See: An orange sky stretches above an azure launchpad. Clustered on or near the pad, colorful balloons of various shapes wait, freshly inflated and straining against the ground. Each balloon is a tight knot of same-colored or distinctly colored cells, full of lift.

Do: Release the balloons. Each one rises into the orange sky — the more air inside it, the higher it floats. A balloon's buoyancy is measured by its fullness: every breath of color it holds carries it one story higher. The launchpad stays bolted to the earth, and the sky fills in behind each balloon as it ascends.

Grid: The sky is 8 columns wide and 8 stories tall. The azure launchpad is the ground floor, permanent and immovable. The orange atmosphere fills every space not claimed by a balloon. Each balloon keeps its shape perfectly as it drifts upward — no stretching, no tumbling, just a clean vertical ascent.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (5306 chars)

We need to infer rule. From example: Input 8x8 grid. Output same size.

Interpretation: orange sky = O, azure launchpad = A at bottom row? In input, bottom row is A's across entire row. That's launchpad, stays same in output (bottom row unchanged). Balloons are clusters of same-colored cells (non-O, non-A) that are "tight knots". They rise upward, moving vertically until they hit something? The sky fills behind each balloon as it ascends (i.e., cells vacated become O). The launchpad stays.

In example1 input: Let's map.

Rows index 0 top to 7 bottom.

Input rows:
0: O O O O O O O O
1: O O O O O O O O
2: O O O O O O O O
3: O O O O O O O O
4: O O O O O O O O
5: O W O O O O O O
6: W W W O O R R R
7: A A A A A A A A

We have colors: W (maroon) cluster at rows5-6 columns0-2? Let's locate.

Row5 col1 = W (since row5: O W O O O O O O) column index1.

Row6: W W W O O R R R (col0,1,2 are W; col5,6,7 are R)

So there are two balloons: one of W (maroon) shape: cells at (5,1), (6,0),(6,1),(6,2). That's a shape maybe 4 cells. Another balloon of R (red) shape: cells at (6,5),(6,6),(6,7) three cells.

They rise upward. In output:

Rows:
0: O O O O O O O O
1: O W O O O O O O
2: W W W O O O O O
3: O O O O O R R R
4: O O O O O O O O
5: O O O O O O O O
6: O O O O O O O O
7: A A A A A A A A

We see W cluster moved up: originally bottom rows5-6 now at rows1-2. Specifically output row1 col1 = W (was row5 col1). Output row2 col0-2 = W (was row6 col0-2). So they moved up by 4 rows? Let's compute shift: original topmost W cell at row5 moved to row1 => shift -4 rows (up 4). Original bottommost W at row6 moved to row2 => also up 4. So whole cluster moved up 4 rows.

R cluster originally at row6 col5-7 moved to row3 col5-7 (output). That's shift -3 rows (up 3). Wait topmost R row6 -> row3 shift -3.

Why different shift amounts? Possibly each balloon rises until it hits either top of grid or another balloon that has already risen? But they rise simultaneously? The rule may be: each balloon rises as far as possible until it would collide with another balloon or the top, but they cannot pass through each other; order maybe based on size (fullness) i.e., more cells rise higher (more buoyancy). The description: "the more air inside it, the higher it floats. A balloon's buoyancy is measured by its fullness: every breath of color it holds carries it one story higher." So number of cells determines how high it goes (i.e., vertical offset). So each balloon moves up by number of cells in its cluster? Let's test: W cluster size = 4 cells. It moved up 4 rows. Yes! R cluster size = 3 cells, moved up 3 rows. That matches.

Thus rule: For each non-O, non-A colored connected component (4-directionally?) treat as balloon. Compute its size (number of cells). Move the entire component upward by that many rows, preserving shape, as long as within grid (no overflow). The launchpad (A row) stays bottom. The vacated cells become O (sky). If moving would go out of bounds (above top), mayb

... [2318 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 705: "Two jousting knights lower their lances and charge, meeting ... [44 chars truncated]" k=2 alts: 1/15 b7249182

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains exactly two colored pixels on a black background. These two pixels share either the same row or the same column, forming an axis between them. They are separated by some distance D (measured cell-to-cell).

Do: 1. Identify the two colored pixels and the axis (row or column) they share. Call them P1 and P2, where P1 is closer to the top or left edge. Let D be the distance between them along the shared axis. 2. Each pixel grows into an open-ended cross shape directed toward the other pixel: a. From P1, extend an arm along the shared axis toward P2, covering (D-3)/2 cells beyond P1 (not counting P1 itself). Call the endpoint of this arm the crossbar position. b. At the crossbar position, draw a perpendicular crossbar centered on the axis, extending 2 cells in each perpendicular direction (5 cells total including the axis cell). c. From the two ends of the crossbar, extend 1 additional cell toward P2 along the shared axis. These form corner pieces. 3. Repeat symmetrically for P2: extend an arm from P2 toward P1 for (D-3)/2 cells, draw a perpendicular 5-cell crossbar at the arm's end, and extend the crossbar endpoints 1 cell toward P1. 4. This leaves a 2-cell gap between the two shapes along the shared axis. The corner extensions from each crossbar face each other across this gap. 5. Each shape is drawn in the color of its originating pixel. The original pixel positions remain colored as part of their respective arms.

Grid: The grid dimensions match the input exactly. The background is black (.). The two colored pixels are always aligned on the same row or same column, with enough surrounding space for the cross shapes to be drawn without clipping the grid edges.

Figurative Descriptions (16 variants)

"Two jousting knights lower their lances and charge, meeting shield-to-shield in the middle of the field."

120b: k=2

See: Two lone knights of different colors stand on a dark tournament field, facing each other across a shared lane — either the same row or the same column.

Do: Each knight charges toward the other, dragging their lance along the shared lane. Halfway to meeting, each knight plants a shield — a broad crossbar perpendicular to their charge — and the edges of the shield curl forward like pauldrons reaching toward the opponent. The two knights never collide; they halt with a narrow no-man's-land between their shields, close enough to lock eyes through the gap. Each knight's trail, shield, and pauldrons are forged in that knight's own heraldic color.

Grid: The tournament field keeps its original dimensions. The dark grass remains untouched except where each knight has carved a colored path. The two knights always share the same row or column, and the field is wide enough that no shield or lance stroke flies off the edge.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (14623 chars)

We need to infer rule. From examples.

Example1: Input grid 10 rows? Actually given 10 rows? Let's count: Input lines show 10 rows? Actually first example input shows 10 rows of 13 columns? Let's count: lines: row0 to row9? They show 10 rows. Input has G at (6,1?) Let's index rows 0-9, columns 0-12. Input: row6: ". G . . . . . . . . . . B". So G at column1, B at column12. So two knights of different colors in same row (row6). Output shows pattern: a cross shape? Let's examine output rows.

Output row4: ". . . . . G G B B . . . ."
Row5: ". . . . . G . . B . . . ."
Row6: ". G G G G G . . B B B B B"
Row7: ". . . . . G . . B . . . ."
Row8: ". . . . . G G B B . . . ."

So looks like knights at same row, they move towards each other, leaving a path (lance) of their color from original position towards center, stopping before meeting, leaving a gap of one cell? Actually they meet at column? Let's compute.

Original positions: G at (6,1). B at (6,12). Distance = 11 cells apart. They move towards each other. In output, G's lance extends from its original position to some point near center. Let's see G cells: row6 columns 0? Actually row6 output: ". G G G G G . . B B B B B". So G occupies columns 1-5? plus column0? Wait column0 is '.'; column1 is G (original). columns2-5 also G. So G occupies columns1-5 inclusive (5 cells). B occupies columns8-12? Actually columns8-12 are B (5 cells). There's a gap at columns6-7? Row6 column6 '.' column7 '.' (gap). So they stopped with two empty cells between? Actually gap of two columns.

Also there are vertical extensions: rows4-8 have G and B in column5? Let's see G vertical line at column5? Row4: columns5? Row4: ". . . . . G G B B . . . ." So columns5=G,6=G,7=B,8=B. Row5: ". . . . . G . . B . . . ." So column5=G, column8=B. Row7 similar. Row8 same as row4.

Thus shape: each knight creates a "cross" shape: a central shield (a perpendicular bar) at the midpoint? The description: each knight charges, dragging lance along lane (row). Halfway to meeting, each knight plants a shield — a broad crossbar perpendicular to their charge — and the edges of the shield curl forward like pauldrons reaching toward the opponent. So the shield is a perpendicular bar (vertical if lane is horizontal) centered at the knight's halfway point? The "edges curl forward" meaning the bar extends forward (toward opponent) maybe like a T shape? Actually output shows for G: vertical line at column5 (the shield) extending up and down from row6 (the lane) and also forward (right) from column5? Wait G's shield is vertical line at column5 rows4-8 (5 cells). That's perpendicular to lane (horizontal). The "edges curl forward" meaning at the ends of the vertical bar (top and bottom) there are horizontal extensions towards opponent? In output, at row4, there is also a G at column6 (right of vertical line) forming a small horizontal extension (pauldron). Similarly at row8 column6 G. So the vertical bar plus rightward extensions at its ends.

Simil

... [11650 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.93)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 706: "A single drop of dye falls into water and stains whatever cl ... [24 chars truncated]" k=1 alts: 9/15 b7256dcd

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input grid has an orange (O) background containing one or more connected components (4-connected) made of magenta (M) cells. Scattered across the grid are single cells of other non-orange, non-magenta colors (e.g., yellow, blue, green, azure, maroon, grey). Each of these colored cells may or may not be 4-adjacent (sharing an edge) to a magenta connected component.

Do: 1. Identify every connected component of magenta (M) cells using 4-connectivity (up, down, left, right). 2. For each non-orange, non-magenta colored cell in the grid, check whether it is 4-adjacent to any cell belonging to a magenta connected component. 3. If the colored cell is adjacent to a magenta component, replace every magenta cell in that component with the colored cell's color. Then replace the colored cell itself with orange (O). 4. If the colored cell is NOT adjacent to any magenta component, simply replace it with orange (O) — it disappears. 5. Any magenta component that has no adjacent colored marker remains magenta. 6. The result is the transformed grid.

Grid: The output grid has the same dimensions as the input grid. The background color is orange (O). All non-orange, non-magenta marker cells from the input are removed (set to orange). Magenta shapes that were painted are replaced entirely by the color of their adjacent marker; unpainted magenta shapes remain as magenta.

Figurative Descriptions (16 variants)

"A single drop of dye falls into water and stains whatever clay sculpture it touches."

120b: k=1

See: An orange lake stretches across the grid. Rising from the water are sculptures made of magenta clay, their cells fused together in connected formations. Floating on the surface nearby are single droplets of vivid color — green, yellow, blue — each one a potent dye waiting to be absorbed.

Do: Each dye droplet reaches out to whatever it touches. If a droplet is brushing against a clay sculpture, the sculpture drinks in the color completely — every cell of connected clay transforms to match the dye. The droplet, spent, dissolves back into the orange water and vanishes. Droplets that float alone, touching no sculpture, simply sink and disappear without a trace. Any sculpture that no droplet touched remains its original magenta clay.

Grid: The lake remains the same size. The orange water fills every space it did before. All the little dye droplets are gone — absorbed or dissolved. Where magenta sculptures once stood, you may now see shapes dyed a new color, or sculptures still wearing their original magenta if no dye ever reached them.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (8116 chars)

We need to infer rule from example. Input grid 5x5? Actually example 1 input 5x5. Let's map.

Input:
Row1: O M M G O
Row2: O O O Y O
Row3: O O O M O
Row4: O O M M M
Row5: O O O M O

Interpretation: O = orange water (lake). M = magenta clay sculptures. G,Y are droplets (green, yellow). Possibly also other colors but only G,Y present.

Output:
Row1: O G G O O
Row2: O O O O O
Row3: O O O Y O
Row4: O O Y Y Y
Row5: O O O Y O

Observations: Original orange O remain same except where changed? In output, many O turned into other colors (G,Y) where M were replaced. Let's see mapping.

Identify connected components of M (magenta). Need to see which are adjacent (4-direction?). Let's label coordinates (r,c). Input M positions:
(1,2),(1,3) adjacent horizontally -> component A size2.
(1,4) is G droplet.
(2,4) Y droplet.
(3,4) M at (3,4) isolated? Adjacent to (4,4) M? Actually (4,4) M, (4,5) M, (4,3) M, (5,4) M. So there is a larger component B of M's at rows 3-5 columns 3-5 maybe.

List M positions:
(1,2),(1,3)
(3,4)
(4,3),(4,4),(4,5)
(5,4)

These are connected? (3,4) adjacent to (4,4) vertically, so part of same component as lower block. Also (4,3) adjacent to (4,4) etc. So component B includes (3,4),(4,3),(4,4),(4,5),(5,4). That's 5 cells.

Component A is (1,2),(1,3) size2.

Now droplets: G at (1,4) adjacent to M? It touches (1,3) left neighbor which is M component A. So G droplet touches component A. According to description, droplet reaches out to whatever it touches; if brushing against a clay sculpture, the sculpture drinks in the color completely — every cell of connected clay transforms to match the dye. Droplet disappears. So component A should become all G. Indeed output rows: (1,2) becomes G, (1,3) becomes G. Yes.

Droplet Y at (2,4) touches which M? Adjacent to (3,4) below which is part of component B. So Y should recolor component B to Y. Output shows component B cells become Y: positions (3,4) Y, (4,3) Y, (4,4) Y, (4,5) Y, (5,4) Y. Indeed output matches.

All other cells become O (orange) unchanged. Droplets themselves disappear (become O). Indeed (1,4) G droplet becomes O in output? Row1 col4 originally G, output row1 col4 O. Yes. (2,4) Y droplet becomes O. Good.

Thus rule: Identify each connected component of M (magenta). If any droplet (any non-O, non-M) of a particular color is orthogonally adjacent to any cell of that component, then recolor entire component to that droplet's color. If multiple droplets of different colors touch same component? Not seen. Probably not happen. If multiple droplets same component maybe same color anyway. If component not touched by any droplet, remains M? In description says remains original magenta. But in output, there were no untouched M. So rule.

Also droplets that are not adjacent to any M become O (disappear). So they vanish.

Now apply to test input grid (7x7). Let's parse.

Rows:

1: Y O M O O O X
2: M M M O O O O
3: O O M O B O O
4: O O O O M M O
5: O O O O M M O
6: M M A O O O O
7: O M O O

... [5170 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.22)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 718: "Each voice in a choir grows louder the more singers join the ... [10 chars truncated]" k=2 alts: 4/15 b91ae062

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: The input is a 3×3 grid where each cell is either black (background) or one of several distinct colors. The number of unique non-black colors present in the grid varies from puzzle to puzzle.

Do: 1. Count the number of unique non-black colors in the input grid. Call this number N. 2. Create an output grid of size (3×N) by (3×N). 3. Replace each cell in the 3×3 input with an N×N block of that cell's color. Black cells become N×N blocks of black. The spatial arrangement of blocks in the output mirrors the arrangement of cells in the input.

Grid: The input is always a 3×3 grid. The output is a (3N)×(3N) grid, where N is the count of unique non-black colors in the input. The background color is black (.). The output is effectively the input scaled up by a factor of N in both dimensions.

Figurative Descriptions (16 variants)

"Each voice in a choir grows louder the more singers join the ensemble."

120b: k=2

See: A small stage with nine positions arranged in a square — some occupied by performers in colored costumes, others left dark and empty. Each performer is a distinct voice in the ensemble.

Do: The more voices present in the choir, the more space each one claims. Every performer swells in presence proportional to the total number of singers — as if each voice resonates with every other, amplifying itself. The empty seats swell too, but remain silent darkness. The chorus arranges itself exactly as they stood on the original stage, just grander.

Grid: The intimate 3×3 rehearsal stage becomes a full concert hall. Each performer's single seat blooms into a square throne whose size is dictated by the richness of the ensemble — a solo act barely grows, but a quintet commands enormous territory. The darkness between performers expands in kind, a respectful silence scaled to match the music.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (3847 chars)

We need to infer rule. Input 3x3, output larger grid. In example1 input:

Row1: . A .
Row2: . M M
Row3: W A .

Output is 9x9? Let's see dimensions: output shown 9 rows each 9 columns. So scaling factor? Possibly each input cell becomes a 3x3 block? Since 3x3 input -> 9x9 output, factor 3. But pattern not simple uniform scaling: For example input cell '.' at (1,1) becomes top-left 3x3 block of '.'? Output rows1-3 cols1-3 are '.' indeed. Input (1,2) = A becomes block of A? Output rows1-3 cols4-6 are A. Yes. Input (1,3) '.' becomes block of '.' at cols7-9 rows1-3. So each cell expands to 3x3 block of its color. But check other cells: Input (2,1) '.' becomes rows4-6 cols1-3 '.'? Output rows4-6 cols1-3 are '.'? Actually rows4-6 cols1-3 are '.'? Let's see output row4: ". . . M M M M M M". So first three are '.' yes. Input (2,2)=M becomes block rows4-6 cols4-6 all M. Indeed row4 cols4-6 M, row5 same, row6 same. Input (2,3)=M becomes block rows4-6 cols7-9 M. Yes. Input (3,1)=W becomes rows7-9 cols1-3 W. Output rows7-9 first three are W. Input (3,2)=A becomes rows7-9 cols4-6 A. Yes. Input (3,3)=. becomes rows7-9 cols7-9 '.'.

Thus rule: each cell expands to a 3x3 block of same color. That's simple scaling factor 3.

Check example2: Input 3x3, output 6x6? Actually output rows: 6 rows, each 6 columns. That's scaling factor 2. Let's verify: Input:

Row1: B . Y
Row2: . Y .
Row3: . B .

Output rows1-2 cols1-2 should be B block? Output row1: "B B . . Y Y". So first two columns B B, matches scaling factor 2. Row1 col3-4 '.' '.' matches input (1,2) '.' block. Row1 col5-6 Y Y matches input (1,3) Y block. So scaling factor equals number of distinct colors? Wait example1 scaling factor 3, example2 scaling factor 2. Maybe scaling factor equals number of non-empty (non-.) cells? Example1 non-'.' cells: A, M, M, W, A =5 distinct? Actually count non-'.' cells =5. Not 3. Maybe scaling factor equals number of unique colors present? Example1 colors: A, M, W => 3 unique colors => scaling factor 3. Example2 colors: B, Y => 2 unique colors (since '.' not counted) => scaling factor 2. That matches! So rule: Determine number of distinct non-'.' colors in input; call that N. Then each cell expands to an N x N block of its color (including '.' becomes block of '.'), preserving arrangement.

Thus for test input, we need to find distinct colors (excluding .). Input:

Row1: . B .
Row2: . A O
Row3: W W .

Colors present: B, A, O, W. That's 4 distinct colors. So scaling factor N=4. Output will be 12x12 (since 3*4 =12). Each input cell becomes 4x4 block.

We need to construct output grid rows 1-12, cols 1-12.

Let's map each input cell to block coordinates:

Input rows i=1..3, cols j=1..3.

Block rows: (i-1)*4+1 to i*4. Block cols: (j-1)*4+1 to j*4.

Now fill with the cell's character.

Let's create.

Input (1,1) = '.' => rows1-4, cols1-4 are '.'.

Input (1,2) = B => rows1-4, cols5-8 are B.

Input (1,3) = '.' => rows1-4, cols9-12 are '.'.

Input (2,1) = '.' =>

... [889 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.03)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 724: "A marching band performs a parade where every third drummer ... [25 chars truncated]" k=1 alts: 13/15 ba26e723

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: A 3-row grid composed entirely of yellow (Y) and black (.) cells. The middle row (row 1) is solid yellow. Rows 0 and 2 alternate between Y and . in a checkerboard pattern, with opposite phases — where row 0 has Y, row 2 has ., and vice versa. The grid width varies but is always odd for rows starting with Y and can be even or odd otherwise.

Do: Some yellow cells are recolored to magenta (M); black cells are never changed. Apply the following rules based on column index (0-based): 1. Middle row (row 1): Replace Y with M at every column that is divisible by 3 (columns 0, 3, 6, 9, ...). 2. For rows 0 and 2, the rule depends on whether the row starts with Y or with .: - If the row starts with Y: Replace Y with M at every column divisible by 6 (columns 0, 6, 12, 18, ...). - If the row starts with .: Replace Y with M at every column where column mod 6 equals 3 (columns 3, 9, 15, 21, ...). All other cells remain unchanged.

Grid: The grid is always exactly 3 rows tall. The width varies across instances. The coordinate system is 0-indexed with row 0 at the top and column 0 at the left. The only colors present are yellow (Y) and black (.), with magenta (M) introduced by the transformation. No cells outside the grid boundary are involved.

Figurative Descriptions (16 variants)

"A marching band performs a parade where every third drummer strikes a magenta cymbal."

120b: k=1

See: A three-lane parade route stretches across the town. The center lane is packed shoulder-to-shoulder with yellow-uniformed marchers. The outer two lanes have marchers spaced out in an alternating pattern — where one lane has a marcher, the other has an empty gap, like two interlocking combs.

Do: A drumbeat sounds every three steps along the center lane, and each marcher standing on that beat swaps their yellow uniform for a magenta one. The outer lanes hear a slower rhythm — a cymbal crash every six steps. But the two outer lanes are offset from each other: whichever lane has a marcher standing at the start of the route hears the crash on the downbeat, while the lane that begins with a gap hears it on the off-beat, halfway between crashes. Wherever a crash lands on an occupied position, that marcher also turns magenta. Empty gaps stay empty — you can't dye what isn't there.

Grid: The parade route is always exactly three lanes wide (top to bottom) but varies in length. The center lane is always full. The top and bottom lanes alternate presence and absence in opposite lockstep. Only yellow marchers can turn magenta; the black gaps between them are just open road.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6017 chars)

We need to decode puzzle. The description: three-lane parade route (grid 3 rows). Center lane always full of Y (yellow). Outer lanes have alternating pattern: one lane has a marcher (Y) then gap (.) then Y etc, the other lane opposite (gap then Y). So pattern like top row: . Y . Y . Y ... bottom row: Y . Y . Y . ... or vice versa.

Transformation: drumbeat every three steps along center lane: each marcher standing on that beat swaps Y to M (magenta). So in center row, positions that are multiples of 3? Starting at step 0? "every three steps along the center lane" likely means positions 0,3,6,...? Let's examine example 1.

Example 1 Input: Let's write rows.

Row1: "Y . Y . Y . Y . Y . Y . Y" (length 13?) Actually count: Y . Y . Y . Y . Y . Y . Y = 13? Let's count: positions 0 Y,1 .,2 Y,3 .,4 Y,5 .,6 Y,7 .,8 Y,9 .,10 Y,11 .,12 Y. So length 13.

Row2: "Y Y Y Y Y Y Y Y Y Y Y Y Y" (13 Y's)

Row3: ". Y . Y . Y . Y . Y . Y ." (13)

Output rows:

Row1: "M . Y . Y . M . Y . Y . M"
Row2: "M Y Y M Y Y M Y Y M Y Y M"
Row3: ". Y . M . Y . Y . M . Y ."

Let's analyze transformation.

Center row (row2) input all Y. Output: M at positions? Output row2: M Y Y M Y Y M Y Y M Y Y M. So M at positions 0,3,6,9,12 (0-index). That's every 3 steps starting at 0. Good.

Now outer rows: Row1 input pattern Y . Y . Y . Y . Y . Y . Y (starting with Y). Output row1: M . Y . Y . M . Y . Y . M. So M at positions 0,6,12? Actually output row1: positions:0 M,1 .,2 Y,3 .,4 Y,5 .,6 M,7 .,8 Y,9 .,10 Y,11 .,12 M. So M at 0,6,12. That's every 6 steps? Wait pattern: crash every six steps, offset depending on lane start. For top lane, it starts with Y (occupied). So it gets crash on downbeat (i.e., at positions 0,6,12...). Indeed M at those positions. For bottom lane (row3) input pattern . Y . Y . Y . Y . Y . Y . (starts with .). Output row3: ". Y . M . Y . Y . M . Y .". M at positions 3,9? Let's see: positions:0 .,1 Y,2 .,3 M,4 .,5 Y,6 .,7 Y,8 .,9 M,10 .,11 Y,12 . . So M at 3 and 9 (and maybe 15 if longer). That's offset by 3 (half of 6) relative to top lane. So rule matches description: crash every 6 steps, lane that starts with a marcher gets crash at positions 0,6,12,... lane that starts with gap gets crash at positions 3,9,15,...

Thus transformation: For each cell that is Y (yellow) in any row, we may turn to M if its column index satisfies certain condition: For middle row: index %3 ==0. For top row: if top row starts with Y (i.e., first cell Y), then index %6 ==0 else index %6 ==3? Actually for top row in example, starts with Y, so M at indices 0,6,12 => index %6 ==0. For bottom row, starts with . so M at indices 3,9 => index %6 ==3.

Thus rule: For outer rows, M if (index %6 == offset) where offset = 0 if cell at index 0 is Y, else offset = 3. Equivalent to (index+ (if start with . then 3 else 0)) %6 ==0? Actually offset = (start with .) ? 3 : 0.

Now test input:

Test Input:
Row1: ". Y . Y . Y . Y . Y . Y . Y . Y ."
Row2: "Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y"
Row3: "Y . Y . 

... [3047 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.94)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 731: "A lock only turns when the right key fits perfectly into its ... [9 chars truncated]" k=1 alts: 0/15 bbb1b8b6

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Example 6

Input

→

Output

Example 7

Input

→

Output

Literal Description

See: The input is a 4-row by 9-column grid. A vertical column of grey (X) cells at column index 4 divides the grid into two 4×4 sub-grids: a left grid and a right grid. The left grid contains a pattern of blue (B) cells and black (.) cells. The right grid contains a pattern of a single non-black color (such as red, magenta, orange, or green) and black (.) cells.

Do: 1. Split the input along the grey (X) column to extract the left 4×4 grid and the right 4×4 grid. 2. Identify all positions that are black (.) in the left grid — these are the 'empty slots.' 3. Identify all positions occupied by the colored (non-black) cells in the right grid. 4. Compare the two sets of positions. If they are exactly equal (same positions, same count), replace every black cell in the left grid with the color from the right grid and output the result. 5. If the positions do NOT exactly match, output the left grid unchanged.

Grid: The input grid is always 4 rows by 9 columns. Column 4 is always grey (X) and serves as a separator. The output grid is always 4×4. The left sub-grid uses only blue (B) and black (.), and the right sub-grid uses only one non-black, non-blue color and black (.). The output grid uses at most two colors: blue and either black or the color from the right grid.

Figurative Descriptions (16 variants)

"A lock only turns when the right key fits perfectly into its grooves."

120b: k=1

See: A grey wall divides two rooms. On the left, a blue vault door has gaps cut into it — empty keyholes waiting. On the right, a colored key floats in darkness, its teeth shaped from a single vivid color.

Do: Slide the key into the lock. If every tooth fits perfectly into every keyhole — no extra teeth, no empty holes left over — the key merges with the vault door, filling each gap with its color. If the key doesn't match the lock, the vault stays sealed and unchanged, its empty keyholes still dark.

Grid: The grey wall always stands in the center, separating lock from key. The vault door is always built from blue steel and dark gaps. The key is always carved from a single color. What emerges is either a door transformed — blue and the key's color fused together — or the same cold, unturned lock.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (4983 chars)

We need to infer rule. From description: grey wall (X) separates left side (vault door) and right side (key). Vault door built from blue (B) with gaps (.)? In example input: grid 4x9? Actually example1 input shows 4 rows, 9 columns. Left side columns before X are vault door region: columns 0-3 are B pattern forming a rectangle with interior maybe? Let's view:

Row0: B B B B X . . . .
Row1: B . . B X . M M .
Row2: B . . B X . M M .
Row3: B B B B X . . . .

So left of X (col0-3) is a blue border rectangle with interior '.' (black) holes. Right of X (col5-8) is key region: colors M (magenta) forming a 2x2 block at rows1-2 cols6-7? Actually columns after X: col5 is '.' (black), col6 M, col7 M, col8 '.'.

Key is a single color (magenta) shape maybe rectangular? The key shape is contiguous region of same color on right side.

The rule: If key shape exactly fits the pattern of holes ('.') in the blue door (i.e., same shape), then fill those holes with key color, merging. Output shows only the left side region (door) with holes filled with M. Output grid is 4x4 (only left side). Indeed output is:

B B B B
B M M B
B M M B
B B B B

So they removed the wall and key region, output only the door region after merging. The key shape matched the holes shape (a 2x2 interior). So they filled interior with M.

If key does not match, output would be unchanged door? Let's infer from test input.

Test Input:

Row0: B B B B X R . . .
Row1: . B B . X R R R R
Row2: . B B . X R . . .
Row3: . . . . X . . . .

Let's parse dimensions: 4 rows, 9 columns.

Left side (cols0-3) door region: columns before X (col4). Let's extract:

Row0 cols0-3: B B B B
Row1: . B B .
Row2: . B B .
Row3: . . . .

So door shape: border maybe? Not full rectangle. Let's view pattern: It seems a 2x2 blue block at rows1-2 cols1-2, with top row full B's, bottom row empty. Actually row0 all B's across 4 columns. Row1 has B at col1,2. Row2 same. Row3 all '.'.

So the door region is not a solid rectangle; maybe the shape of door is defined by B cells (blue). The gaps are '.' cells within the bounding rectangle? In example, door region was a solid rectangle of B border with interior '.' holes. Here door region maybe shape of B cells forms something like a "U" shape? Let's map coordinates (r,c):

B cells:
(0,0)(0,1)(0,2)(0,3)
(1,1)(1,2)
(2,1)(2,2)

That's like a top row full, then a 2x2 block below offset one column in.

The gaps ('.') within the convex hull? Not sure.

Key region: columns after X (col5-8). Let's extract:

Row0 col5-8: R . . .
Row1: R R R R
Row2: R . . .
Row3: . . . .

Key color is R (red). Shape: positions of R:

Row0: (0,5)
Row1: (1,5)(1,6)(1,7)(1,8)
Row2: (2,5)

So shape is like a T? Actually coordinates relative to key region (col5 is first after X). Let's shift to key local coordinates (col5->0). Then shape cells: (0,0), (1,0)(1,1)(1,2)(1,3), (2,0). That's like an upside-down L? Not matching door holes.

We need to determine if key shape matches the door's gaps. What are the gaps? In

... [1986 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 733: "Two distant shores reflected in a shrinking lake, merging as ... [23 chars truncated]" k=1 alts: 5/15 bc1d5164

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: The input is a 5-row by 7-column grid with a single non-black color. Colored cells appear only in the two leftmost columns (0 and 1) and the two rightmost columns (5 and 6). The middle three columns (2, 3, 4) are always black. The grid has a vertical axis of symmetry in its structure: the left pair of columns (0–1) and the right pair (5–6) encode two overlapping copies of a 3-column pattern, sharing their inner columns (1 and 5). Row 2 (the middle row) is always entirely black, creating a similar horizontal split: rows 0, 1 and rows 3, 4 encode two overlapping copies of a 3-row pattern sharing their inner rows (1–3).

Do: Produce a 3×3 output grid by combining the left and right column pairs and compressing the five rows into three, using an OR-merge (a cell is colored if either of its source cells is colored, otherwise black): 1. Output column 0 comes from input column 0. 2. Output column 1 is the OR-merge of input column 1 and input column 5 (colored if either is colored). 3. Output column 2 comes from input column 6. For rows: 1. Output row 0 comes from input row 0. 2. Output row 1 is the OR-merge of input rows 1, 2, and 3 (colored if any of those rows has a colored cell in that column). 3. Output row 2 comes from input row 4. Apply both column and row mappings simultaneously to determine each of the 9 output cells.

Grid: The input grid is always 5 rows by 7 columns. The background is black (.). Exactly one non-black color is used per puzzle. Columns 2, 3, and 4 are always entirely black, serving as a separator between the left column pair (0–1) and the right column pair (5–6). The output grid is always 3 rows by 3 columns, using the same single color on a black background.

Figurative Descriptions (16 variants)

"Two distant shores reflected in a shrinking lake, merging as the water drains away."

120b: k=1

See: A wide landscape stretches across seven columns, but the heart of it—three columns in the center—is nothing but still, dark water. Life clings only to the narrow shores on the far left and far right, two columns wide on each side. A similar emptiness runs through the middle row, a quiet channel dividing north from south. The colored marks are like campfires lit on these shores, mirrored signals calling to each other across the void.

Do: The dark water evaporates and the land folds inward like a closing book. The left shore and the right shore slide together until they overlap, their inner edges merging into a single shared column. Where either shore had a campfire burning, the merged land keeps the flame—darkness only wins where both sides were dark. The same folding happens vertically: the northern and southern banks collapse toward the center channel, and wherever any of those middle rows carried a signal, the fire persists in the compressed land.

Grid: The wide world is five rows by seven columns, with a single color of campfire against the black night. The three central columns are always empty water. After the fold, only a small three-by-three island remains—the essence of both shores unified, fires combined, darkness only where silence reigned on every side.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6215 chars)

We need to understand the puzzle. The description: grid 5 rows x 7 columns. Central three columns are empty (black). The leftmost two columns and rightmost two columns may have colored cells (single color per grid? In examples, only one color appears). The transformation: fold horizontally (left onto right) merging inner edges (column 3? Actually central three columns are empty, so after folding left two columns onto right two columns, they overlap into a single column? Let's parse.

We have 7 columns: indices 0-6. Central three columns are columns 2,3,4 (0-index). Those are always empty (black). Left shore: columns 0,1. Right shore: columns 5,6.

Folding: left shore slides right, right shore slides left, overlapping. The inner edges (columns 1 and 5) merge into a single shared column (maybe column 2 after folding). The outer edges (columns 0 and 6) become outer columns of final 3x3 grid.

Similarly vertical folding: rows 0-4 (5 rows). Middle row is row 2 (0-index). That's empty (black). Top shore rows 0,1. Bottom shore rows 3,4. Fold vertically: top rows slide down, bottom rows slide up, merging inner edges (rows 1 and 3) into a single row (row 1 after folding). Outer edges rows 0 and 4 become outer rows of final 3x3.

Thus final grid is 3x3: rows correspond to top outer, middle merged, bottom outer; columns correspond to left outer, middle merged, right outer.

Now rule for colors: "Where either shore had a campfire burning, the merged land keeps the flame—darkness only wins where both sides were dark." Means after folding, for each overlapping cell (i.e., where left and right overlapped, or top and bottom overlapped), we take OR of colors: if either side has a colored cell (non-black), that color appears. But what if both sides have colors, possibly different? In examples, only one color appears overall, so not needed. But we need to decide tie-breaking: maybe if both have colors, they are same? Could be that they are same color always. We'll assume they are same or we can pick any? Might be that they never conflict.

Also for non-overlapping cells (outer edges), they just stay as they were (since only one side). So final 3x3 grid is composed of:

- Top row: from original row 0 (leftmost column 0) becomes top-left, row0 col0; top-middle from overlapping of row0 col1 (left inner) and row0 col5 (right inner?) Wait need mapping.

Let's map coordinates.

Original grid indices: rows 0-4, cols 0-6.

After folding horizontally: columns mapping to final 3 columns:

- Final column 0 (left outer) corresponds to original column 0 (leftmost) after folding (since right side outer column 6 also maps to final column 2 maybe). Actually final grid width 3: left outer, middle merged, right outer.

Left outer final column = original column 0 (leftmost). Right outer final column = original column 6 (rightmost). Middle final column = result of merging original column 1 (left inner) and column 5 (right inner). Since central columns 2,3,4 are empty and disappe

... [3239 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.22)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 740: "A corner fortress wall collapses inward, leaving rubble laye ... [46 chars truncated]" k=3 alts: 0/15 bd5af378

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid contains exactly two colors. One color (the border color) forms an L-shape consisting of one complete row and one complete column that share a single corner cell (the L-corner). The other color (the fill color) occupies the remaining rectangular region. The L-corner is always at one of the four corners of the grid.

Do: 1. Identify the border color (the L-shaped row+column) and the fill color (the rectangle). Identify the L-corner where the border row and column meet. 2. Layer 0 (outermost): Replace every cell in the border row and border column with the fill color, then set the L-corner cell to azure (A). 3. Layer 1 (one step inward): On the row adjacent to the border row (toward the grid center) and the column adjacent to the border column (toward the grid center), write the border color into all cells that are not already part of Layer 0. Then set the corner cell where this inner row and inner column meet to azure (A). 4. Interior: Fill all remaining cells (those not in Layer 0 or Layer 1) with azure (A). Then draw a diagonal of the fill color starting from the interior corner nearest the L-corner and extending one cell per step toward the diagonally opposite corner of the interior, for min(interior_height, interior_width) cells.

Grid: The output grid has the same dimensions as the input grid. It uses the same two colors from the input plus azure (A). The grid can be any rectangular shape; the L-shaped border may appear in any of the four corner orientations (top-left, top-right, bottom-left, bottom-right).

Figurative Descriptions (16 variants)

"A corner fortress wall collapses inward, leaving rubble layers and a crack running through the foundation."

120b: k=3

See: A fortress stands at one corner of the land. Its walls form an L-shaped rampart — one wall running the full length of a border row and another running the full length of a border column, meeting at a watchtower in the corner. The interior courtyard is filled with a single color of stone.

Do: The fortress is struck and its walls collapse inward. The outermost rampart crumbles: the wall stones transform into courtyard stone, and the watchtower corner is replaced by azure rubble. One step inward, the fallen wall leaves a shadow — a single row and column of the original wall color mark where the debris landed, with another piece of azure rubble at their intersection. Beyond this debris line, azure dust fills the entire interior — except for a deep diagonal crack of courtyard stone that splits the foundation, running from the corner nearest the old watchtower outward toward the opposite corner, one cell per step, as far as the smaller dimension of the interior allows.

Grid: The land keeps its original shape. The ruins are built from three materials: the old wall stone, the courtyard stone, and azure rubble. The fortress watchtower may have stood at any of the four corners of the land.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (14564 chars)

We need to decode rule.

We have description: fortress L-shaped rampart along border row and column meeting at corner (watchtower). Interior courtyard filled with single color. So input grid: one corner has a different color? Let's examine examples.

Example1 Input grid 5x6? Actually lines: "W W W W W O" repeated 4 rows, then "O O O O O O". So dimensions 5 rows? Wait count rows: first four rows are "W W W W W O" (6 columns). Then fifth row "O O O O O O". So 5 rows, 6 columns. The L-shaped rampart: walls are W (maroon) along top row? Actually top row is W W W W W O: first five cells W, last O. Left column? Let's view grid: column0 all rows? Row0 col0 W, row1 col0 W, row2 col0 W, row3 col0 W, row4 col0 O. So left column has W for first four rows, then O at bottom row. Bottom row is all O. Rightmost column? column5 values: rows0-3 O, row4 O. So the L shape maybe along top row (full length) and left column (full length) meeting at top-left corner? But top row has W except last cell O. So top row length 5 W then O. Left column has W for rows0-3, then O at row4. So L shape of W along top row (first 5 cells) and left column (first 4 rows). The interior courtyard is O (orange) filling rest of grid (including bottom row and rightmost column). So indeed fortress at top-left corner.

Output example1: Let's see transformation. Output grid same size 5x6. Provided output rows:
Row0: A W A A O W
Row1: A A W A O W
Row2: A A A W O W
Row3: O O O O A W
Row4: W W W W W A

Interpretation: Azure (A) appears where walls collapsed? The outermost rampart crumbles: wall stones become courtyard stone (O) and watchtower corner replaced by azure (A). In output, the original W walls (top row and left column) have been transformed: The outermost wall (the original W) becomes A? Actually look at top row: original W W W W W O. Output top row: A W A A O W. So first cell (corner) becomes A (azure). The next cells? original W at col1 becomes W still? Actually output col1 is W (unchanged). col2 original W becomes A. col3 original W becomes A. col4 original O becomes O (unchanged). col5 original O becomes W (now W). Hmm.

Let's map rule more concretely: The description says: outermost rampart crumbles: wall stones transform into courtyard stone, and watchtower corner replaced by azure rubble. One step inward, the fallen wall leaves a shadow — a single row and column of the original wall color mark where the debris landed, with another piece of azure rubble at their intersection. Beyond this debris line, azure dust fills the entire interior — except for a deep diagonal crack of courtyard stone that splits the foundation, running from the corner nearest the old watchtower outward toward the opposite corner, one cell per step, as far as the smaller dimension of the interior allows.

So we need to produce output with three materials: old wall color (W), courtyard stone (O), azure (A). The interior after transformation is azure except a diagonal of courtyard stone.

Let's test

... [11600 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.07)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 741: "Russian nesting dolls rotate their painted shells inward" k=1 alts: 11/15 bda2d7a6

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid consists of concentric rectangular rings, each filled with a single color. The outermost ring is one color, the next ring inward is a second color, the next is a third, and so on. There are exactly three distinct colors used across all rings. If there are more than three rings, the colors cycle, so the fourth ring reuses the first color, the fifth reuses the second, and so on.

Do: 1. Identify the concentric rectangular rings from outermost to innermost, noting the color of each ring. 2. List the distinct colors in the order they first appear going inward, producing a sequence [C1, C2, C3]. 3. Apply a cyclic shift backward (rotate left by one position): C1 maps to C3, C2 maps to C1, and C3 maps to C2. In other words, each color is replaced by the one that preceded it in the sequence, with the first color wrapping around to become the last. 4. Rebuild the grid with the same ring structure but with each ring's color replaced according to this mapping. The grid dimensions and ring geometry remain unchanged.

Grid: The output grid has the same dimensions as the input grid. The structure of concentric rectangular rings is preserved exactly — only the color assigned to each ring changes. The rings are axis-aligned rectangles, uniformly one cell thick, nested from the border inward to the center.

Figurative Descriptions (16 variants)

"Russian nesting dolls rotate their painted shells inward"

120b: k=1

See: The grid is a set of nested picture frames, each painted a different color, stacked like matryoshka dolls from the outer wall to the hidden center. Three paint colors take turns dressing each successive frame as they spiral inward.

Do: The dolls play a game of musical chairs: when the music stops, each doll puts on the coat of the doll that was standing just inside it. The outermost doll inherits the innermost color, as if the smallest doll's secret heart has been turned inside out to become the new skin of the world.

Grid: The frame structure is sacred — no frame moves, shrinks, or grows. Only the paint on each frame changes, as though every shell has been carefully unscrewed, the colors passed one neighbor inward, and the whole thing reassembled in place.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (9591 chars)

We need to infer rule. Input is a grid with nested frames. Example 1: Input grid 8x8. Let's write coordinates. Input outer border is A (azure). Next inner border (one cell inside) is . (black) except corners? Actually row2: A . . . . . . A. So second row has A at ends, interior . . . . . . . So inner border is . (black). Next inner border (third row) has A . X X X X . A. So third row: first A, then ., then X X X X, then ., then A. So third layer: X (grey) interior rectangle. Then fourth row: A . X A A X . A. So inside that, there is A region (azure) inside X region. So pattern: nested frames of colors: outermost A, then . , then X, then A, then inner core maybe .? Actually inner core is 2x2 of A? Let's see rows 4-5 columns 4-5 are A A. So inner core is A. So frames: colors sequence from outer to inner: A, ., X, A (inner core). Output grid: colors shifted? Output shows outermost border now X (previous third layer). Let's examine output: first row all X. So outermost becomes X (previous third layer). Second row: X A A A A A A X. So second row border X, interior A's. So second layer becomes A (previous outermost). Third row: X A . . . . A X. So third layer becomes . (previous second layer). Fourth row: X A . X X . A X. So fourth layer becomes X (previous third layer). Fifth row same. Sixth row: X A . . . . A X. Seventh row: X A A A A A A X. Eighth row all X.

Thus rule: rotate colors of frames inward? Actually each frame takes color of the next inner frame (the one just inside it). The innermost frame takes color of outermost? Let's see innermost (center 2x2) in output: rows4-5 cols4-5 are X X (from previous third layer). In input innermost was A. So innermost got color of previous third layer (one step outward). So seems each frame takes color of the frame just outside it? Wait outermost took color of third layer (two steps inward). Let's map layers indices: layer0 outermost, layer1 next, layer2 next, layer3 innermost (core). Input colors: L0=A, L1=., L2=X, L3=A. Output colors: L0=X (was L2), L1=A (was L0), L2=. (was L1), L3=X (was L2). So it's a cyclic shift of colors forward by one layer? Actually mapping: each layer gets color of layer+2? Let's see shift: L0 gets L2, L1 gets L0, L2 gets L1, L3 gets L2. That's like each layer takes color of the layer just inside it (i.e., inner neighbor) but outermost takes innermost? Wait L0 takes L1? No L0 took L2 not L1. So maybe they rotate three colors taking turns? The description: "Three paint colors take turns dressing each successive frame as they spiral inward." So there are three colors rotating. In example, colors involved are A, ., X (three colors). They rotate inward: each frame gets color of the next inner frame, but the innermost gets color of outermost? Let's test: If we rotate inward (each frame takes color of inner neighbor), then L0 would get L1 (.), but output L0 is X, not . So not that.

If rotate outward (each frame takes color of outer neighbor), then L0 would get L-1 (none) maybe innermos

... [6750 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 784: "Gatekeepers on opposite walls signal to the hallway guard be ... [11 chars truncated]" k=1 alts: 3/15 cbded52d

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The grid is divided into a 3×3 arrangement of 2×2 blocks, separated by single rows and columns of black (.) cells acting as dividers. Each 2×2 block is filled with blue (B) by default. Some blocks contain one or more non-blue colored cells at specific positions within their 2×2 area — these are 'marked' blocks. The marked blocks appear along the edges of the 3×3 arrangement (the top row, bottom row, left column, and right column of blocks). The center block and other interior-edge blocks may start as plain all-blue blocks.

Do: For each row of blocks (top, middle, bottom) and each column of blocks (left, center, right), examine the two edge blocks that bookend that row or column. If both edge blocks in a row share a non-blue color at the same position within their 2×2 block, copy that color into the corresponding position of the middle block in that row. Similarly, if both edge blocks in a column share a non-blue color at the same position within their 2×2 block, copy that color into the corresponding position of the middle block in that column. If only one edge block has a non-blue color at a given position (and the opposite edge block does not), the middle block is not affected at that position. Apply all such propagations to produce the output.

Grid: The grid is 8 columns × 8 rows. It consists of a 3×3 arrangement of 2×2 content blocks. The blocks are separated by single-cell-wide dividers filled with black (.) — specifically, rows 2 and 5 are entirely black, and columns 2 and 5 are entirely black. The 2×2 blocks occupy the remaining cells. Colors used are blue (B) as the default fill, with up to a few non-blue accent colors (such as orange, green, yellow, red, azure, magenta, maroon) appearing in marked blocks.

Figurative Descriptions (16 variants)

"Gatekeepers on opposite walls signal to the hallway guard between them."

120b: k=1

See: A walled city is laid out as a 3×3 grid of guard towers, separated by dark empty corridors. Most towers are staffed entirely by blue-uniformed soldiers. A few towers along the city walls have a special officer — someone wearing a distinctive colored sash — standing at a particular post within the tower.

Do: When two towers on opposite walls of the same street both have an officer wearing the same colored sash at the same post, the tower in the middle of that street recruits a matching officer for that post — as if the two sentries are signaling across the corridor, and the middle tower answers the call. A lone officer with no counterpart on the far wall sends no signal, and the middle tower ignores them.

Grid: The city is 8×8 cobblestones. Nine 2×2 guard towers sit in a 3×3 arrangement, divided by single-width dark corridors (rows 2 and 5, columns 2 and 5). Blue is the standard uniform. Colored sashes — orange, green, yellow, red, and others — mark the special officers.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (11962 chars)

We need to infer rule. From description: 8x8 grid with 3x3 towers each 2x2 separated by corridors rows 2 and5 (0-index?). Actually rows 0-7. Corridors are rows 2 and5, columns 2 and5 (single width). So towers are blocks: top-left rows0-1 cols0-1, top-middle rows0-1 cols3-4, top-right rows0-1 cols6-7, middle-left rows3-4 cols0-1, etc.

Within each tower, most cells are blue (B). Some cells have colored sashes (non-blue). The rule: when two towers on opposite walls of same street both have an officer wearing same colored sash at same post (position within 2x2), then middle tower gets that colored sash at that same position. So we need to look at each "street": horizontal street: left and right towers (west/east) opposite each other across middle column corridor? Actually street is row corridor? Let's think: towers on opposite walls of same street: For horizontal street, towers left and right of middle column (i.e., same row block) across the vertical corridor column 2? Wait towers are arranged 3x3. Streets are rows and columns of corridors. So there are vertical corridors (columns 2 and5) separating left-middle-right towers. Horizontal corridors (rows 2 and5) separating top-middle-bottom towers.

Thus for each pair of towers opposite each other across a corridor, we compare corresponding positions (relative within 2x2). If both have same non-blue color at same relative position, then the tower in middle (the one adjacent to both across the corridor) gets that color at that position.

Specifically: For vertical street: left tower (col block 0-1) and right tower (col block 6-7) across middle tower (col block 3-4). If both left and right have same colored sash at same relative cell (e.g., top-left of their 2x2), then middle tower gets that color at that same relative cell.

Similarly for horizontal street: top tower (rows0-1) and bottom tower (rows6-7) across middle tower (rows3-4) for each column block.

Also there are diagonal? No.

Now check Example1 to verify.

Example1 Input grid (I'll map). Let's write towers.

Rows0-1, cols0-1: top-left tower:
Row0 col0 B, col1 B
Row1 col0 B, col1 G (green at position (1,1) relative bottom-right)

Top-middle tower rows0-1 cols3-4:
Row0 col3 O (orange) at (0,0) relative? Actually within tower coordinates (0,0) top-left is col3 row0. Row0 col4 B
Row1 col3 B, col4 B

Top-right tower rows0-1 cols6-7:
Row0 col6 B, col7 B
Row1 col6 B, col7 G (green at bottom-right)

Middle-left tower rows3-4 cols0-1:
Rows3-4: row3 col0 B, col1 B; row4 col0 B, col1 B (all blue)

Middle-middle tower rows3-4 cols3-4:
row3 col3 B, col4 B; row4 col3 B, col4 B (all blue)

Middle-right tower rows3-4 cols6-7:
row3 col6 B, col7 B; row4 col6 B, col7 B (all blue)

Bottom-left tower rows6-7 cols0-1:
row6 col0 B, col1 B; row7 col0 B, col1 B (all blue)

Bottom-middle tower rows6-7 cols3-4:
row6 col3 O, col4 B; row7 col3 B, col4 B

Bottom-right tower rows6-7 cols6-7:
row6 col6 B, col7 B; row7 col6 B, col7 G

Now output: they changed some cells.

Ou

... [8977 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.02)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 786: "A fabric swatch dreams of becoming the whole bolt of cloth." k=1 alts: 6/15 ccd554ac

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Example 6

Input

→

Output

Literal Description

See: The input is a small grid of N rows by M columns containing a pattern made of one color and black (background) cells. The grid functions as a single tile or motif.

Do: Tile the input grid to fill a larger output grid by repeating it M times horizontally and N times vertically. That is, place copies of the input side by side to form a row of M tiles, then stack N such rows on top of each other. The result is an (N×N) by (M×M) output grid where the input pattern repeats seamlessly in both directions.

Grid: The input grid has dimensions N rows by M columns. The output grid has dimensions (N×N) rows by (M×M) columns. The number of horizontal repetitions equals M (the column count of the input) and the number of vertical repetitions equals N (the row count of the input). Colors are preserved exactly; no cells change color.

Figurative Descriptions (16 variants)

"A fabric swatch dreams of becoming the whole bolt of cloth."

120b: k=1

See: A small swatch of fabric displays a motif woven from a single dye color against an undyed background. This fragment remembers the dimensions of the loom it was cut from.

Do: The swatch is placed on the loom and allowed to replicate itself — it knows how many times to repeat across the warp by counting its own columns, and how many times to stack along the weft by counting its own rows. The pattern propagates like a crystal growing from a seed, each copy seamlessly kissing its neighbor at the edges until the bolt is complete.

Grid: The seed swatch carries its own multiplication instructions in its shape. A wider swatch stretches further across; a taller swatch grows further down. The dye never bleeds or shifts — every copy is a perfect echo of the original.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (5484 chars)

We need to infer rule. Input is a small pattern; output is larger tiled pattern replicating based on its dimensions. In example 1, input grid size? Let's parse.

Example 1 Input (3 rows, 3 columns):
Row1: . . A
Row2: A A A
Row3: A . .

Output appears 9x9? Let's see: Output rows count: they show 9 rows, each with 9 entries (space separated). Indeed output is 9x9. So they tiled the 3x3 pattern 3 times across and 3 times down (i.e., repeat both dimensions). But not simple repetition; pattern seems interleaved? Let's examine output first row: ". . A . . A . . A". That's three copies of row1 concatenated horizontally. Row2: "A A A A A A A A A" = three copies of row2. Row3: "A . . A . . A . ." = three copies of row3. So first three rows correspond to first copy row1-3. Then rows 4-6 repeat same three rows again? Let's see row4: ". . A . . A . . A" same as row1. Row5: "A A A A A A A A A" same as row2. Row6: "A . . A . . A . ." same as row3. Row7-9 repeat again. So output is just tiling the input pattern 3x3 times (both dimensions). So rule: replicate input pattern across both axes as many times as its own size (i.e., repeat N times where N = number of rows = number of columns? Actually input is square? In example, 3x3, repeated 3 times each direction, resulting size = 3*3 =9. So repeat count equals dimension size (rows count). If non-square? Might be repeat rows count times vertically and columns count times horizontally. So output dimensions = rows*rows by cols*cols? But if rows != cols, maybe repeat rows times vertically and columns times horizontally. Let's test with test input: size? Provided test input grid: appears 5 rows, each 5 columns? Let's list:

Row1: . A A . .
Row2: A A A A A
Row3: . A A . .
Row4: A A A A A
Row5: . A A . A

That's 5x5. So we would repeat 5 times horizontally and vertically, resulting 25x25 grid. That's huge. But maybe they expect that. In ARC tasks, output often is larger but manageable. 25x25 is okay.

Thus output will be each input row repeated 5 times horizontally to make a long row, and then that block repeated 5 times vertically.

So we need to produce 25 rows each of 25 characters.

Let's compute.

First, define input rows as strings of characters (with spaces). We'll produce concatenated horizontally 5 times.

Row1 pattern: ". A A . ."
Concatenate 5 times: ". A A . . . A A . . . A A . . . A A . . . A A . . . A A . ."? Wait need to ensure spaces between cells. Actually each row is sequence of 5 cells. Repeating 5 times yields 25 cells: pattern repeated.

Let's write row1 repeated 5 times:

Sequence of 5 cells: [.,A,A,.,.]
Repeat 5 times: indices 1-25.

Let's write as characters separated by spaces.

I'll produce row1_out = ". A A . . . A A . . . A A . . . A A . . . A A . . . A A . ."

Count: each repeat adds 5 cells, total 25. Let's count groups: we have 5 repeats, each ". A A . ." So row1_out = ". A A . . . A A . . . A A . . . A A . . . A A . . . A A . ."

Now row2 input: "A A A A A" (all A). Repeating 5 times yields

... [2484 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.04)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 792: "Two armies join forces across a golden bridge to form a unit ... [9 chars truncated]" k=3 alts: 12/15 ce4f8723

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input consists of two 4×4 grids stacked vertically, separated by a horizontal row of yellow (Y) cells. The top grid contains blue (B) and black (.) cells. The bottom grid contains red (R) and black (.) cells. The yellow row always spans the full width and acts as a divider.

Do: 1. Split the input at the yellow divider row to obtain the top grid (blue/black) and the bottom grid (red/black). 2. Create an output grid of the same dimensions as either sub-grid. 3. For each cell position: if the top grid has blue OR the bottom grid has red at that position, set the output cell to green (G). Otherwise, set it to black (.). This is a logical OR of the two grids' non-black cells.

Grid: The input grid is 4 columns wide and 9 rows tall (4-row top grid + 1 yellow divider row + 4-row bottom grid). The output grid is 4×4. Background color is black (.). The only colors in the output are green (G) and black (.).

Figurative Descriptions (16 variants)

"Two armies join forces across a golden bridge to form a united front."

120b: k=3

See: Two kingdoms sit on opposite sides of a golden bridge. The northern kingdom flies blue banners from some of its towers, while the southern kingdom lights red fires atop others. The darkness between them marks empty watchtowers where no one stands guard.

Do: The golden bridge opens and the two kingdoms merge their defenses. Any watchtower that had a blue banner flying OR a red fire burning sends a soldier to the unified army, marked in green. Only towers that were dark and abandoned in BOTH kingdoms remain empty in the new realm.

Grid: The world is a tall fortress split by a golden bridge spanning the full width. Above and below the bridge lie matching castle walls. The united army's banner—green against the darkness—flies only where at least one kingdom once kept watch.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (6197 chars)

We need to infer rule. From description: two kingdoms north and south separated by golden bridge (maybe a row of Y's? In examples, there is a row of Y Y Y Y in middle). Input grids have top part (north) with B's and . (black) maybe representing blue banners and empty towers. Bottom part (south) with R's and . representing red fires and empty. The Y row is the bridge. Output only includes top part? In examples output only shows top part rows (same number as top part?). Let's examine Example1.

Input dimensions: Let's count rows. Provided as lines:

B B . .
B . B .
B B . B
B B B B
Y Y Y Y
R R . R
. . R .
. R . .
R . R .

That's 9 rows, each 4 columns. The Y row is row5 (index 4). So top part rows 1-4 (north), bottom part rows 6-9 (south). Output:

G G . G
G . G .
G G . G
G G G G

That's 4 rows, 4 columns. So output corresponds to top part rows (north) transformed: each cell becomes G if either north cell was B (blue) OR corresponding south cell (same column) was R (red). If both are '.' (black) then output '.'? Let's test.

We need mapping: For each column, row position? The rule likely merges the two halves: For each position (i,j) in north half (rows above Y), look at north cell (maybe B or .) and the corresponding south cell (same column, but row mirrored? The description: "Only towers that were dark and abandoned in BOTH kingdoms remain empty". So if both north and south have '.' (black), output '.'; else output G (green). So output is green wherever at least one of the two corresponding cells (north and south) is not '.' (i.e., B or R). Let's verify with Example1.

Create arrays:

North rows (1-4):
1: B B . .
2: B . B .
3: B B . B
4: B B B B

South rows (6-9):
6: R R . R
7: . . R .
8: . R . .
9: R . R .

Now compute per column per row (same row index? The description says "merge their defenses" maybe aligning rows: row1 north aligns with row1 south? But there are 4 north rows and 4 south rows, same count. So row1 north aligns with row1 south (row6). Let's test.

Row1 north: B B . .
Row1 south: R R . R
Combine: any B or R => G, else . So positions:
col1: B or R => G
col2: B or R => G
col3: . and . => . (both .) => .
col4: . and R => G
Result row1: G G . G matches output row1. Good.

Row2 north: B . B .
Row2 south: . . R .
Combine: col1 B or . => G; col2 . and . => . ; col3 B or R => G ; col4 . and . => . => G . G . matches output row2: G . G . correct.

Row3 north: B B . B
Row3 south: . R . . (row8)
Combine: col1 B or . => G ; col2 B or R => G ; col3 . and . => . ; col4 B or . => G => G G . G matches output row3: G G . G correct.

Row4 north: B B B B
Row4 south: R . R . (row9)
Combine: col1 B or R => G ; col2 B or . => G ; col3 B or R => G ; col4 B or . => G => G G G G matches output row4. Good.

Thus rule confirmed.

Now Example2.

Input rows count? Let's list.

Rows:
1 B B B .
2 . B . B
3 . . B B
4 B B . B
5 Y Y Y Y
6 . . . R
7 . . . R
8 R R R R
9 R R . R

North rows 1-4, South ro

... [3440 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.11)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 798: "Three stained glass windows shine through each other, and th ... [23 chars truncated]" k=2 alts: 7/15 cf98881b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: The input is a 4×14 grid containing three 4×4 panels separated by two vertical columns of red (R) cells. The left panel contains yellow (Y) cells and black (.) cells. The middle panel contains maroon (W) cells and black (.) cells. The right panel contains blue (B) cells and black (.) cells. Each panel represents a layer of colored positions against a black background.

Do: Extract the three 4×4 panels by removing the two red divider columns. Overlay the three panels into a single 4×4 output grid using the following priority: yellow (Y) takes highest priority, then maroon (W), then blue (B), then black (.). For each cell position, check the left panel first — if it contains Y, the output is Y. If the left panel is black at that position, check the middle panel — if it contains W, the output is W. If both the left and middle panels are black, check the right panel — if it contains B, the output is B. If all three panels are black at that position, the output is black (.).

Grid: The input grid is 4 rows by 14 columns. It consists of three 4×4 content panels separated by two 4×1 columns of red (R) cells at columns 4 and 9. The output grid is 4×4. Colors used are yellow (Y), maroon (W), blue (B), red (R) as dividers, and black (.) as background.

Figurative Descriptions (16 variants)

"Three stained glass windows shine through each other, and the brightest light wins."

120b: k=2

See: A triptych of stained glass windows is mounted in a cathedral wall, separated by pillars of red stone. The left window glows with golden sunlight (Y), the center with a deep wine-colored warmth (W), and the right with a cool blue moonlight (B). Where the glass is clear (.), no light passes through at all.

Do: Remove the red stone pillars and stack the three windows one behind the other, letting light shine through. Golden sunlight is the most powerful — wherever it streams through the left window, it dominates. Where the golden pane is clear, the wine-colored warmth from the middle window seeps through instead. Only where both the gold and wine panes are clear can the blue moonlight from the farthest window reach the viewer. Where all three windows are clear glass, only darkness remains.

Grid: The cathedral wall holds three square panes of equal size, braced apart by two narrow columns of red stone. Each pane is a mosaic of colored and clear tiles. When the pillars are removed and the panes are layered, the congregation sees a single window where the nearest light always outshines the farther ones.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (7507 chars)

We need to infer rule. The description: three windows side by side separated by red pillars (R columns). Input grid contains three panes separated by columns of R? Let's examine example1 input dimensions. Input lines: appears 4 rows? Actually given as space-separated? Let's reconstruct.

Example1 Input lines (maybe 4 rows, each row length 15?). Let's write rows:

Row1: . . . Y R . . . W R . . . .
Row2: Y Y . Y R W . W . R . . . .
Row3: Y . Y Y R . W W . R B B . B
Row4: . Y Y Y R . W . . R B B B B

So 4 rows, 15 columns. There are two red columns (R) acting as pillars separating three windows each of width? Let's locate R columns: column indices (0-based). Row1: positions: 0 .,1 .,2 .,3 Y,4 R,5 .,6 .,7 .,8 W,9 R,10 .,11 .,12 .,13 .,14 ?. Actually row1 length 15, last char maybe . . Let's count: given ". . . Y R . . . W R . . . ." that's 15 entries: indices 0-14. So R at index4 and index9. So windows: columns 0-3 (left window width 4), columns5-8 (center window width 4), columns10-14 (right window width5?). Wait after second R at index9, remaining columns 10-14 =5 columns. But maybe right window also width 4 but extra column? Let's check other rows: Row2: "Y Y . Y R W . W . R . . . ." also 15 entries, R at 4 and 9. After second R, columns10-14 =5 entries: . . . . ?. Actually row2 after R at 9, we have ". . . ."? Let's list: Row2 entries: 0 Y,1 Y,2 .,3 Y,4 R,5 W,6 .,7 W,8 .,9 R,10 .,11 .,12 .,13 .,14 ?. Wait there are 15 entries, last maybe . . Let's count: given "Y Y . Y R W . W . R . . . ." that's 15: after R at 9, we have . (10), . (11), . (12), . (13), . (14). So right window width 5 columns of all '.' in this row.

Row3: "Y . Y Y R . W W . R B B . B" 15 entries, R at 4 and 9. Right side columns10-14: B B . B (that's 4 entries) but we need 5 columns; maybe column10 B,11 B,12 .,13 B,14 ??? Actually there are 5 entries after index9: B (10), B (11), . (12), B (13). That's only 4. Wait maybe missing one at end? The row ends with B, so maybe there are 5 entries: B (10), B (11), . (12), B (13), (nothing). Let's recount: row3 tokens: Y(0) .(1) Y(2) Y(3) R(4) .(5) W(6) W(7) .(8) R(9) B(10) B(11) .(12) B(13). That's 14 tokens, not 15. Maybe there is missing token at end? The given row: "Y . Y Y R . W W . R B B . B" that's 14 tokens. So maybe the grid is 4x14? Let's check row1 tokens count: ". . . Y R . . . W R . . . ." that's 14? Count: .0, .1, .2, Y3, R4, .5, .6, .7, W8, R9, .10, .11, .12, .13. That's 14 tokens. So grid width 14. Good. So windows: left width 4 (cols0-3), middle width 4 (5-8), right width 4 (10-13). The red pillars at col4 and col9 separate. So each pane is 4x4.

Thus transformation: we have three 4x4 windows. Need to stack them: nearest (left) overrides, then middle, then right, with priority: Y > W > B > .? Actually description: golden (Y) most powerful, where gold present dominates. Where gold clear (i.e., .) then wine (W) from middle shows. Only where both gold and wine clear can blue (B) from farthest show. Where all three 

... [4633 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 806: "Shooting stars streak across a night sky from where they ign ... [4 chars truncated]" k=2 alts: 13/15 d13f3404

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: A 3×3 grid with a black (.) background containing a small number of non-black colored cells scattered at various positions. Each non-black cell represents a 'seed' that will generate a diagonal streak in the output.

Do: Create an output grid of 6×3 (6 rows, 6 columns — double the input dimensions). Initialize all cells to black. For every non-black cell at position (row, col) in the input, place that color repeatedly along a diagonal streak: at (row, col), (row+1, col+1), (row+2, col+2), and so on, continuing until the position falls outside the output grid boundaries. If two streaks overlap at the same position, the cell retains whichever color is placed there (in practice, streaks from distinct input cells do not collide).

Grid: The input is always a 3×3 grid. The output is a 6×6 grid (exactly double the input dimensions in both rows and columns). The background color is black (.). The coordinate system is zero-indexed with (0,0) at the top-left corner, rows increasing downward, and columns increasing to the right.

Figurative Descriptions (16 variants)

"Shooting stars streak across a night sky from where they ignite."

120b: k=2

See: A small patch of night sky holds a few bright stars scattered among the darkness. Each star burns with its own unique color — these are meteors caught at the moment of ignition, frozen just before they begin their descent.

Do: The sky expands as night deepens, and each meteor finally falls. Every star streaks diagonally downward and to the right, leaving a trail of its own light — an afterimage that repeats its color at each step of the descent. The meteors fall in parallel, each one tracing its own luminous scar from wherever it first appeared, tumbling until it passes beyond the edge of the visible sky.

Grid: The initial sky is a small square window. As the meteor shower unfolds, the visible sky doubles in both height and width to accommodate the full length of every falling streak. The void of space remains black wherever no star has burned its path.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (3850 chars)

We need to infer rule. From examples: Input is 3x3? Example1 input 3x3. Output 6x6 (double size). The rule: each non-black cell (colored) creates a diagonal line down-right across the expanded grid, repeating its color each step, starting at its original position scaled? Let's examine.

Example1 input grid coordinates (row, col) 0-index.

Input:
row0: . Y .
row1: . A .
row2: R . .

So colors: Y at (0,1), A at (1,1), R at (2,0).

Output 6x6. Let's map.

Output rows (0-5) columns (0-5):
Row0: . Y . . . .
Row1: . A Y . . .
Row2: R . A Y . .
Row3: . R . A Y .
Row4: . . R . A Y
Row5: . . . R . A

We can see diagonal lines: Y appears at (0,1), (1,2), (2,3), (3,4), (4,5). That's a diagonal down-right from its original position (0,1) extending to edge. A appears at (1,1), (2,2), (3,3), (4,4), (5,5). R appears at (2,0), (3,1), (4,2), (5,3). So each colored cell creates a diagonal line down-right, including its original cell, extending until hitting bottom or right edge of output grid. Output grid size is double input dimensions (3->6). Why double? Because the longest diagonal from top-leftmost cell may need extra space? Actually if we just extend diagonally within same size, some lines would go out of bounds. Doubling ensures enough space for all lines to fit.

Specifically, output size = input size *2. The original positions are placed at same coordinates? Let's check: Input (0,1) Y appears at output (0,1) same. Input (1,1) A appears at output (1,1) same. Input (2,0) R appears at output (2,0) same. So original grid placed at top-left corner of output (same coordinates). Then for each colored cell, we fill cells (r + k, c + k) for k=0..(N*2 - max(r,c) -1) maybe until hitting edge.

Thus rule: For each non-'.' cell, copy its color along the diagonal down-right across the expanded grid.

Now Example2: Input 3x3:
row0: M B .
row1: G . .
row2: . . .

Colors: M (0,0), B (0,1), G (1,0).

Output 6x6:
Row0: M B . . . .
Row1: G M B . . .
Row2: . G M B . .
Row3: . . G M B .
Row4: . . . G M B
Row5: . . . . G M

Check diagonals: M at (0,0) appears at (0,0),(1,1),(2,2),(3,3),(4,4),(5,5). B at (0,1) appears at (0,1),(1,2),(2,3),(3,4),(4,5). G at (1,0) appears at (1,0),(2,1),(3,2),(4,3),(5,4). Works.

Thus rule confirmed.

Now test input 3x3:
row0: . . G
row1: . . .
row2: . Y W

Coordinates: G at (0,2). Y at (2,1). W at (2,2). Also maybe other colors? That's all.

We need output 6x6.

Place original cells at same positions.

Now extend each diagonal down-right.

First G at (0,2). Diagonal positions: (0,2), (1,3), (2,4), (3,5). Continue until out of bounds (col 6 would be out). So positions: (0,2) G, (1,3) G, (2,4) G, (3,5) G.

Second Y at (2,1). Positions: (2,1), (3,2), (4,3), (5,4). Also (6,5) out. So Y at those.

Third W at (2,2). Positions: (2,2), (3,3), (4,4), (5,5).

Now we need to fill rest with '.'.

Let's construct 6x6 grid rows 0-5, cols 0-5.

Initialize all '.'.

Place G line:
(0,2)=G
(1,3)=G
(2,4)=G
(3,5)=G

Place Y line:
(2,1)=Y
(3,2)=Y
(4,3)=Y
(5,4)=Y

P

... [871 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 813: "Yellow lanterns illuminate the corridor between them, reveal ... [40 chars truncated]" k=3 alts: 7/15 d2acf2cb

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains cells of black (.), maroon (M), and yellow (Y). Yellow cells appear in pairs that share either the same row or the same column. Between each Y pair, the intervening cells are composed of M and . values. No orange (O) or azure (A) cells appear in the input.

Do: 1. Locate all yellow (Y) cells in the grid. 2. Match each Y cell with its partner: the other Y cell that lies in the same row or same column. Each Y belongs to exactly one pair. 3. For each Y pair, identify every cell strictly between them along their shared row or column. 4. Transform each intervening cell: replace maroon (M) with orange (O), and replace black (.) with azure (A). 5. Leave all other cells unchanged, including the Y cells themselves.

Grid: The grid dimensions vary across instances. The grid uses a sparse palette: black (.) as background, maroon (M) as structural fill, and yellow (Y) as boundary markers. Y cells always occur in pairs aligned along a single row or column. The output grid has the same dimensions as the input, with orange (O) and azure (A) appearing only in the segments between Y pairs.

Figurative Descriptions (16 variants)

"Yellow lanterns illuminate the corridor between them, revealing what the maroon shadows were hiding."

120b: k=3

See: A dark landscape of shadow (black) and maroon brick walls, with pairs of yellow lanterns mounted at key positions. Each lantern has a twin somewhere along its row or column, but the corridor between them sits unlit — you can only see the bricks and darkness filling the gap.

Do: When two lanterns face each other, their combined glow floods the corridor between them. The light transforms everything it touches: maroon bricks glow orange in the warmth, and the empty darkness brightens to azure, like a night sky at dawn. The lanterns themselves remain steady yellow — they are the source, not the subject, of the change.

Grid: The world is built from black emptiness and maroon masonry, with yellow lanterns always placed in facing pairs along a shared line. Outside the lanterns' shared gaze, everything stays as it was. Only the space caught between twin flames is transfigured — orange where brick stood, azure where nothing was.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (12043 chars)

We need to infer rule. From description: black '.' and maroon 'W'? Wait color key: . = black, B=blue, R=red, G=green, Y=yellow, X=grey, M=magenta, O=orange, A=azure, W=maroon. In examples, we see only ., M, Y, O, A, etc. No W appears. So maroon is W but not used. In inputs, we have '.' (black), M (magenta), Y (yellow), O (orange), A (azure). In outputs, some cells change: In example1, input row2: Y O A O A A A A Y becomes output row2: Y M . M . . . . Y. So many changes. Let's analyze.

We need to identify lanterns: yellow Y are sources. They appear in pairs along same row or column with only '.' (black) or maybe other colors between? The description: pairs of yellow lanterns mounted at key positions. Each lantern has a twin somewhere along its row or column, but the corridor between them sits unlit — you can only see the bricks and darkness filling the gap. When two lanterns face each other, their combined glow floods the corridor between them. The light transforms everything it touches: maroon bricks glow orange, and the empty darkness brightens to azure. Lanterns remain Y.

Thus rule: For each pair of Y in same row or column with no other Y between them? The cells between them (exclusive) are transformed: any M (magenta) becomes O (orange) (since maroon bricks? Wait maroon is W, but we have M as magenta representing bricks? In description maroon bricks become orange. In grid, maroon is W but not present. However M is magenta maybe representing bricks. In examples, M cells between Y become O? Let's check.

Example1: Input row2: Y O A O A A A A Y. Between the two Y's are O A O A A A A (positions 2-8). Output row2: Y M . M . . . . Y. So O turned to M? A turned to .? Not matching description.

Let's examine column wise maybe.

Better to map colors: In description, maroon bricks (W) become orange (O). Empty darkness (.) becomes azure (A). So transformation: W -> O, . -> A. But we have M (magenta) maybe representing bricks? In examples, M appears a lot as background. Maybe M corresponds to maroon bricks? But color key says M=magenta, W=maroon. However maybe they used M as maroon bricks (since W not used). Let's test: In example1, between Y's there were O and A originally, not M. So maybe O and A are the transformed colors already? Wait input row2 had O and A already. Output row2 changed them to M and . . So transformation seems opposite: O -> M, A -> . (black). That suggests that the corridor between Y's is being "cleared" to something else? Let's examine other rows.

Example1, row5 input: Y A O O O O A A Y. Output row5: Y . M M M M . . Y. So between Y's, A O O O O A A become . M M M M . . . So A turned to . (black), O turned to M (magenta). So indeed rule: within corridor, O becomes M, A becomes . (black). So O (orange) becomes M (magenta) and A (azure) becomes . (black). That seems like shifting colors outward? Let's see other examples.

Example2: Input row6: Y . . M M M M . Y. Output row6: Y A A O O O O A Y. So between Y's, . . M

... [9160 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.33)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 815: "Each blue star is a compass rose, casting its colored light ... [32 chars truncated]" k=1 alts: 15/15 d364b489

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Literal Description

See: The input is a grid with a black (.) background containing several blue (B) cells scattered at various positions. No other colors are present.

Do: For each blue (B) cell in the grid, add four colored neighbors in the cardinal directions: place red (R) one cell above, orange (O) one cell to the left, magenta (M) one cell to the right, and azure (A) one cell below. If any of these positions fall outside the grid boundaries, skip that marker. The blue cell itself remains unchanged. All other cells stay black.

Grid: The grid is 10x10 with a black (.) background. Row 0 is the top and column 0 is the left. The grid dimensions remain the same between input and output.

Figurative Descriptions (16 variants)

"Each blue star is a compass rose, casting its colored light in the four cardinal directions."

120b: k=1

See: A dark night sky dotted with bright blue stars, each one alone in the void, quietly waiting to orient lost travelers.

Do: When dawn begins, each blue star awakens as a compass rose, projecting its cardinal beams into the surrounding darkness — a warm glow to the north, a sunset hue to the west, a violet flash to the east, and a cool shimmer to the south. The star itself remains blue at the center of its cross. Stars near the edge of the sky simply lose whichever beam would fly beyond the horizon.

Grid: The night sky is a square canvas of darkness. North is at the top, west is at the left. The sky does not grow or shrink when the stars ignite their beacons.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6425 chars)

We need to infer rule. From example: Input grid 10x10 with B's at certain positions. Output has B's remain at same positions? Let's compare.

Input example1 coordinates (row,col 0-index):
Row0 col5 B
Row2 col0 B
Row3 col9 B
Row5 col5 B
Row8 col2 B
Row9 col9 B

Output grid: Let's map.

Row0: ". . . . O B M . . ."
So at (0,5) B remains B. Additional O at col4, M at col6. So left of B got O (west), right got M (east). Also maybe north/south? Row0 has no north (top), so no north beam (R). South? At row1 col5? Output row1 col5 is '.'? Row1: "R . . . . A . . . ." So at (1,0) R, (1,5) A. Wait A at col5? Actually row1 col5 is A (since columns 0-9). So south of B (row0 col5) is row1 col5 = A. So south beam is A (cool shimmer). West beam is O (warm glow). East beam is M (violet flash). North beam would be out of bounds, so omitted.

Now check other B's.

B at (2,0). Output row2: "B M . . . . . . . R"
So at (2,0) B stays. East beam at (2,1) M. South beam at (3,0) A? Wait row3 col0 is A? Row3: "A . . . . . . . O B". Yes row3 col0 A. West beam would be out of bounds, omitted. North beam at (1,0) R (since row1 col0 is R). Indeed row1 col0 is R. So north beam appears at (1,0) R. Good.

B at (3,9). Output row3 col9 B. West beam at (3,8) O? Row3 col8 O yes. East out of bounds. South beam at (4,9) B? Wait row4 col9 is B? Row4: " . . . . . R . . . A". Actually row4 col9 is A. So south beam A at (4,9). North beam at (2,9) R? Row2 col9 is R yes.

B at (5,5). Output row5: ". . . . O B M . . ."
So same pattern: west O at (5,4), east M at (5,6), north R at (4,5)?? Row4 col5 is R yes. South A at (6,5) row6 col5 is A? Row6: ". . . . . A . . . ." yes.

B at (8,2). Output row8: ". O B M . . . . . R"
So west O at (8,1), east M at (8,3), north R at (7,2) row7 col2 is R? Row7: ". . R . . . . . . ." yes. South A at (9,2) row9 col2 is A? Row9: ". . A . . . . . O B" yes A at col2.

B at (9,9). Output row9 col9 B. West O at (9,8) O yes. North R at (8,9) R yes. South out of bounds, east out.

Thus rule: For each blue cell B, place colored beams in four cardinal directions: north = R (red), west = O (orange), east = M (magenta), south = A (azure). If direction would go out of grid, omit. Also beams may overlap? Not in example but could. Also original B stays unchanged.

Thus transformation: For each B, set cell north (row-1, same col) to R unless already something? Possibly if multiple B's set same cell, maybe priority? Not needed.

Now apply to test input.

Test Input grid 10x10:

Let's list B positions.

Row0: ". B . . . . . . . B" => B at (0,1) and (0,9)
Row1: all . none
Row2: ". . . . . . . B . ." => B at (2,7)
Row3: ". . . B . . . . . ." => B at (3,3)
Row4: none
Row5: none
Row6: "B . . . . . . B . ." => B at (6,0) and (6,7)
Row7: none
Row8: none
Row9: ". . . . B . . . . ." => B at (9,4)

Now we need to produce output grid same size, start with copy of input? In example output, other cells that were '.' remain '.' unless overwritten by beams. So start with '.' eve

... [3440 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.99)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 817: "A strand of holiday lights where every other bulb switches o ... [9 chars truncated]" k=1 alts: 14/15 d406998b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: A grid with exactly 3 rows. Each column contains exactly one grey (X) cell; all other cells are black (.). The grey cells trace a zigzag path across the columns.

Do: Determine the index of the last column (width minus 1) and note whether it is even or odd. For every grey (X) cell in the grid, check the parity of its column index. If the column index shares the same parity (even/odd) as the last column index, recolor that cell green (G). Otherwise, leave it as grey (X). All black (.) cells remain unchanged.

Grid: The grid is always exactly 3 rows tall. Width varies but each column contains exactly one non-background cell. The background color is black (.). The only non-background colors are grey (X) in the input, and grey (X) plus green (G) in the output.

Figurative Descriptions (16 variants)

"A strand of holiday lights where every other bulb switches on at dusk"

120b: k=1

See: A single strand of holiday lights is strung across three hooks, zigzagging up and down. Each vertical post holds exactly one unlit grey bulb; the rest of the strand is dark wire.

Do: As evening falls, the strand checks which side of the street the last post stands on. Bulbs that stand on the same side as that final post flicker to life in green. The others stay dormant and grey, waiting for a different night.

Grid: The strand always spans exactly three hooks tall. The dark wire is black (.), unlit bulbs are grey (X), and bulbs that catch the evening current glow green (G).

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10946 chars)

We need to infer rule. From description: three hooks tall (grid height = 3). Strand is a path zigzagging across three hooks? Input shows black '.' and grey 'X' positions. Output changes some X to G based on side of final post? Need to interpret.

Look at Example1. Input grid 3 rows, columns? Let's write coordinates. I'll parse.

Input rows:
Row0: ". . X . . X . X . . . X ."
Let's count: tokens separated by space. Row0: [.,.,X,.,.,X,.,X,.,.,.,X,.] length 13? Let's count: 1 .,2 .,3 X,4 .,5 .,6 X,7 .,8 X,9 .,10 .,11 .,12 X,13 . => 13 columns.

Row1: "X . . . X . X . . X . . X" => length 13.

Row2: ". X . X . . . . X . X . ." => length 13.

Output rows:
Row0: ". . G . . X . X . . . X ."
Row1: "G . . . G . G . . X . . G"
Row2: ". X . X . . . . G . G . ."

So rule: some X become G. Which ones? Let's map positions.

Create coordinates (r,c). Input X positions:
Row0 col2,5,7,11? Wait col indices starting 0.
Row0: col2 X, col5 X, col7 X, col11 X.
Row1: col0 X, col4 X, col6 X, col9 X, col12 X.
Row2: col1 X, col3 X, col8 X, col10 X.

Output: G positions:
Row0 col2 becomes G (was X). Others unchanged: col5 stays X, col7 X, col11 X unchanged.
Row1: col0 becomes G (was X). col4 becomes G (was X). col6 becomes G (was X). col9 stays X. col12 becomes G (was X).
Row2: col8 becomes G (was X). col10 becomes G (was X). Others unchanged.

So which X become G? It seems maybe those that are part of a continuous diagonal "strand" that goes from top-left to bottom-right? Let's visualize pattern: maybe the strand is a path that moves down-right or up-right each step, like a zigzag across three rows. Starting at some column, goes through rows 0-2 repeatedly. The X's represent bulbs (grey). The strand is the set of X's that are connected in a zigzag pattern (adjacent horizontally? maybe diagonal?). The rule: turn those X that are on same side as final post? Need to interpret description: "single strand of holiday lights is strung across three hooks, zigzagging up and down. Each vertical post holds exactly one unlit grey bulb; the rest of the strand is dark wire." So each column (vertical post) has exactly one X (bulb) among three rows. Indeed check input: each column should have exactly one X? Let's verify column wise for example1. Column0: row0 ., row1 X, row2 . => one X. Column1: row0 ., row1 ., row2 X => one X. Column2: row0 X, row1 ., row2 . => one X. Column3: row0 ., row1 ., row2 X? Wait column3 row2 X yes. So each column has exactly one X. Good.

Thus the strand is the sequence of X's across columns, forming a path that moves up/down each column (maybe staying same row? but each column only one X). The path zigzags: row index changes each column, moving up or down by 1? Let's check example1 sequence of rows per column (0-12):
col0 row1
col1 row2
col2 row0
col3 row2
col4 row1
col5 row0
col6 row1
col7 row0
col8 row2
col9 row1
col10 row2
col11 row0
col12 row1

Now see transitions: 1->2 (down), 2->0 (up two?), that's not just +/-1. Actually rows are

... [8072 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.82)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 818: "The emerald hearth draws travelers home by the roads they al ... [11 chars truncated]" k=1 alts: 4/15 d43fd935

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains exactly one 2x2 block of green (G) cells, which serves as a central anchor. Scattered across the grid are individual colored dots (non-green, non-black single cells). Some of these dots share a row or column with one or more cells of the green block; others do not.

Do: For each non-green colored dot in the grid, check whether it lies on the same row or the same column as any cell of the 2x2 green block. If it does, draw a line of that dot's color from the dot toward the nearest edge of the green block, filling every empty (black) cell between the dot and the adjacent edge of the green block. The dot itself and the green block cells remain unchanged. If a dot does not share a row or column with any green block cell, leave it unchanged. All other cells remain as they are.

Grid: The grid is 10x10 with a black (.) background. The 2x2 green block can appear anywhere in the grid. Colored dots are single cells scattered at various positions. The grid uses a 0-indexed coordinate system with row 0 at the top and column 0 at the left.

Figurative Descriptions (16 variants)

"The emerald hearth draws travelers home by the roads they already walk."

120b: k=1

See: A glowing emerald hearth (the 2×2 green block) sits in the darkness, while lonely travelers of various colors are scattered across the land. Some travelers happen to stand on a road—a shared row or column—that leads straight to the hearth; others wander paths that never cross it.

Do: Each traveler who finds themselves on a road aligned with the hearth feels its warmth and leaves a trail of their own color behind them as they are pulled toward it—like footprints glowing in the snow. The trail stretches from where the traveler stands all the way to the hearth's doorstep, but the traveler never moves and the hearth is never disturbed. Those who share no road with the hearth remain where they are, unaware of its light.

Grid: The land is a 10×10 expanse of night (black). The emerald hearth is a 2×2 square of green warmth. The travelers are single colored sparks scattered across the darkness, each a different hue marking which road they would illuminate.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (11581 chars)

We need to infer rule from example. Input grid 10x10 with black '.' default. There's a 2x2 green block (G) somewhere. In example1, locate G block: rows 3-4 (0-index? Let's index rows 0-9). Input example1 rows:

Row0: ". O . . . . A . . ."
Row1: ". . . . . . . . . ."
Row2: ". . . O . G G . . A"
Row3: ". . . . . G G . . ."
Row4: ". . . . . . . . . ."
Row5: ". A . . . . . . . ."
Row6: ". . . . . . . . A ."
Row7: ". . . A . . . . . ."
Row8: ". O . . . . . . . ."
Row9: ". . . . . O . . . O"

Green block at (row2,col5),(row2,col6),(row3,col5),(row3,col6). That's 2x2.

Now output: see changes. Let's compare.

Output rows:

Row0 unchanged: ". O . . . . A . . ."
Row1: ". . . . . . A . . ." (A added at col6? Actually row1 col6 originally '.' now 'A')
Row2: ". . . O O G G A A A"
Row3: ". . . . . G G . . ."
Row4: ". . . . . O . . . ."
Row5: ". A . . . O . . . ."
Row6: ". . . . . O . . A ."
Row7: ". . . A . O . . . ."
Row8: ". O . . . O . . . ."
Row9: ". . . . . O . . . . O"

Observations: The "travelers" are colored cells other than G and O? Actually there are A (azure) and O (orange) in input. In output, many O's appear forming lines connecting those colored cells to the green block? Let's see.

In input, there are A at (0,6), (2,9), (5,1), (6,8), (7,3). O at (0,1), (2,3), (8,1), (9,5), (9,9). Also maybe others.

The rule: For each traveler (non-green) that shares a row or column with any cell of the green block, we fill the cells between them (including traveler? maybe not) with that traveler's color, forming a line segment to the nearest green cell? In output, we see O's added along columns/rows aligning with green block.

Let's test: Green block rows 2-3, cols5-6.

Check traveler A at (0,6). Same column 6 as green block (col6). So line from (0,6) down to green block row2? In output, row1 col6 becomes A (added). Row2 col6 already G (green). So they added A at row1 col6 (the cell between traveler and block). Not at row0 col6 (traveler stays). So rule: fill cells between traveler and block, exclusive of both ends? Actually they fill all cells between traveler and the block, not including traveler (original stays) and not overwriting green. So they added A at row1 col6.

Traveler A at (2,9). Same row 2 as green block (row2). Columns 5-6 are green. So line from col9 leftwards to col6? In output row2: after original O at col3? Wait row2 output: ". . . O O G G A A A". Let's map columns 0-9: col0 '.',1'.',2'.',3'O',4'O',5'G',6'G',7'A',8'A',9'A'. Input row2 had O at col3, G at 5,6, A at col9. So they added O at col4 (between O at col3 and green block? Actually O at col3 shares row with green block, so fill between col3 and col5? That would be col4 becomes O. Indeed output col4 O. Also they added A's at col7,8 (between A at col9 and green block at col6). So fill cells between traveler and nearest green cell along row.

Traveler A at (5,1). Same column? column1 not aligning with green block columns 5-6, rows 2-3. Not same row or column, so no change. Indeed output 

... [8626 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.02)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 819: "Every tribe has a sacred sigil, and scattering the warriors ... [66 chars truncated]" k=3 alts: 15/15 d4469b4b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Example 6

Input

→

Output

Example 7

Input

→

Output

Literal Description

See: A 5x5 grid containing scattered cells of a single non-black color (blue, red, or green) against a black (.) background. The specific spatial arrangement of the colored cells varies but is irrelevant to the transformation. The only thing that matters is which color is present.

Do: Identify which color appears in the input grid. Produce a 3x3 output grid based solely on the color identity: (1) If the color is blue (B), output a plus/cross shape: row 1 = . X ., row 2 = X X X, row 3 = . X . (2) If the color is red (R), output an inverted-T shape: row 1 = X X X, row 2 = . X ., row 3 = . X . (3) If the color is green (G), output a bottom-right L shape: row 1 = . . X, row 2 = . . X, row 3 = X X X. Each output shape contains exactly 5 grey (X) cells and 4 black (.) cells.

Grid: Input: a 5x5 grid with cells that are either black (.) or one of three possible colors (blue B, red R, or green G). Only one non-black color appears per grid. Output: a 3x3 grid using grey (X) for the pattern and black (.) for background.

Figurative Descriptions (16 variants)

"Every tribe has a sacred sigil, and scattering the warriors across the battlefield doesn't change the banner they march under."

120b: k=3

See: A battlefield (5×5) is strewn with warriors of a single tribe — blue, red, or green — scattered among empty darkness. Their positions are chaotic, as if flung by a storm, but every warrior wears the same color, and that color is the name of their tribe.

Do: Ignore where the warriors fell — their arrangement is a feint. Instead, read the color of their armor to learn which tribe has taken the field. Each tribe carries an ancestral sigil, a compact grey emblem stamped onto a 3×3 seal. The blue tribe's sigil radiates outward from the center in four directions. The red tribe's sigil stands tall, crowned with a broad brim. The green tribe's sigil hunkers into one corner, bracing two walls that meet at the foundation. Summon the correct sigil and the battle resolves.

Grid: The battlefield is a 5×5 expanse of darkness dotted with a single tribe's colored warriors. The ancestral seal is a 3×3 tablet where the sigil is rendered in grey (X) against black emptiness — always five stones placed, four spaces left bare.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (2736 chars)

We need to infer rule. Input 5x5 with only one color (B,R,G) plus . (black). Output is 3x3 with X's forming a pattern based on tribe.

We have three patterns described metaphorically:

- Blue tribe's sigil radiates outward from the center in four directions. That suggests a plus shape: center plus up/down/left/right. That's 5 cells (center + 4). In output example 1 (blue), output grid:
. X .
X X X
. X .
That's a plus shape indeed. So blue => plus.

- Red tribe's sigil stands tall, crowned with a broad brim. That suggests maybe a T shape? "stands tall" vertical line, "crowned with a broad brim" horizontal line at top. That's a T shape (5 cells). Output example 2 (red):
X X X
. X .
. X .
That's a T shape (top row three X, middle column down). Yes.

- Green tribe's sigil hunkers into one corner, bracing two walls that meet at the foundation. That suggests an L shape (corner). Output example 3 (green):
. . X
. . X
X X X
That's an L shape rotated? Let's see: bottom row full three X, rightmost column top two X, forming an L shape with corner at bottom right? Actually shape: cells (row3,col1-3) all X, plus (row1,col3) and (row2,col3). That's like a reversed L (like a corner with vertical line on right side and horizontal line at bottom). That's an L shape (5 cells). So green => L shape (corner) with the corner at bottom right? But orientation may be determined by something else? In example, green pattern appears oriented with corner at bottom left? Wait coordinates: 3x3 grid rows 1-3, columns 1-3. Output:
Row1: . . X (col3)
Row2: . . X (col3)
Row3: X X X (row3 all)
So shape is like an L rotated 90 degrees clockwise? Actually typical L shape (like corner) could be vertical line on left and horizontal bottom. Here vertical line on right, horizontal bottom. So corner at bottom right? The corner cell is bottom right (row3,col3) which is part of both lines. So orientation is bottom right corner.

Thus green => L shape anchored at bottom right.

Now we need to determine which tribe is present in test input. Input grid includes G and . only? Let's see test input:

Row1: . G . G G
Row2: . . G . .
Row3: G . . . .
Row4: . . G . G
Row5: . . . . G

Only G present, no B or R. So tribe is green. So output should be the green sigil: L shape anchored at bottom right (as per example). However maybe orientation could depend on something else like distribution of G's? But description says ignore arrangement, just color determines pattern. So output is fixed L shape.

Thus output 3x3 grid with X's as described.

Let's verify pattern: Example 3 matches L shape bottom right. So we output same.

Thus answer grid:

Row1: . . X
Row2: . . X
Row3: X X X

Using '.' for black, 'X' for grey.

Provide as block with spaces.

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.22)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 872: "A blue ball bouncing between the walls of a hallway" k=1 alts: 7/15 e179c5f4

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a rectangular grid filled entirely with black (.) cells except for a single blue (B) cell located at the bottom-left corner (last row, first column). The grid has a fixed height of 10 rows and a variable width.

Do: Replace every black cell with azure (A). Then trace a diagonal bouncing path for the blue (B) cell starting from the bottom-left corner (last row, column 0), moving upward one row at a time. The B cell moves diagonally toward the right wall: each row up, the column index increases by 1. When the column reaches the rightmost position (width − 1), the direction reverses and the column decreases by 1 each row. When the column reaches 0, it reverses again, bouncing back to the right. Continue this zigzag until every row has exactly one B cell. Specifically, for each row from bottom (row H−1) to top (row 0), compute the column as follows: take the row's distance from the bottom (d = H−1 − row_index), compute d modulo (2 × (width − 1)), and if the result is less than or equal to width − 1 the column equals the result, otherwise the column equals 2 × (width − 1) minus the result. Place B at that column and A everywhere else in that row.

Grid: The output grid has the same dimensions as the input (same number of rows and columns). The background color is azure (A). Each row contains exactly one blue (B) cell. The grid uses a coordinate system where row 0 is the top and the last row is the bottom; column 0 is the left edge.

Figurative Descriptions (16 variants)

"A blue ball bouncing between the walls of a hallway"

120b: k=1

See: A dark, empty hallway stretches upward, ten stories tall. At the very bottom-left corner, a blue ball rests on the floor, ready to be launched.

Do: The lights come on — the darkness everywhere turns to a cool azure glow. The blue ball is kicked upward and to the right, ricocheting off the walls like a bouncing DVD screensaver logo. Each floor it rises, it drifts one step sideways. When it hits the right wall, it bounces back toward the left; when it hits the left wall, it bounces right again. It leaves its blue mark on every floor it passes through, painting exactly one footprint per story as it zigzags its way to the top.

Grid: The hallway keeps its original shape. The azure light fills every tile except where the ball has bounced — those tiles stay blue. Every floor has exactly one blue scuff mark. The ball starts its journey from the bottom-left corner and works its way upward.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (4072 chars)

We need to infer rule. Input is a grid with '.' black everywhere except a single B at bottom-left corner (row bottom, column left). Output in example 1 is a 10x3 grid (same dimensions as input). Output contains A (azure) and B (blue) placed in pattern. Let's examine pattern.

Input 10 rows, 3 columns. B at row10 (bottom), col1. Output rows correspond? Let's list output rows (top to bottom):

Row1: A B A
Row2: B A A
Row3: A B A
Row4: A A B
Row5: A B A
Row6: B A A
Row7: A B A
Row8: A A B
Row9: A B A
Row10: B A A

We need to see pattern of B's: positions of B in each row.

Row1 col2
Row2 col1
Row3 col2
Row4 col3
Row5 col2
Row6 col1
Row7 col2
Row8 col3
Row9 col2
Row10 col1

So B moves like a bouncing ball moving up rows, shifting sideways each row, bouncing off walls (col1 and col3). Starting at bottom-left (row10 col1). Then moves up one row to row9, shift right one column => col2 (B). Row8 shift right => col3 (B). Row7 would shift right but hits wall, so bounce left => col2 (B). Row6 shift left => col1 (B). Row5 shift left would hit wall, bounce right => col2 (B). Row4 shift right => col3 (B). Row3 shift right would bounce left => col2 (B). Row2 shift left => col1 (B). Row1 shift left would bounce right => col2 (B). That matches pattern.

All other cells become A (azure). So rule: fill all cells with A except the path of the moving B which retains B. The path is a diagonal bounce pattern starting from the B's original location (bottom-left) moving upward each row, shifting horizontally by +1 each step, bouncing off left/right edges.

Now test input: grid 10 rows, 5 columns (since given 5 columns). B at bottom-left (row10 col1). Need to produce output 10x5 with A everywhere except B's path.

We need to simulate bounce across width 5 columns.

Columns indices 1..5.

Start: row10 col1 (B). Direction: moving up each row, horizontal direction initially to the right (+1). At each step (row-1), we attempt to move horizontally by current direction. If next column would be out of bounds (0 or 6), we reverse direction before moving? In example, when at col3 (rightmost) with direction +1, next would be col4 out of bounds, so bounce: direction becomes -1, then move to col2. That matches pattern.

Similarly at leftmost col1 with direction -1, bounce to +1 then move to col2.

Thus we can simulate.

Let's simulate rows from bottom (row10) up to row1.

Initialize col =1, dir = +1.

We'll record B positions per row.

Row10: col1 (given). After marking, move to next row (row9): attempt col+dir =2 (within 1..5) so no bounce. col=2. dir stays +1.

Row9: col2. Next row8: col+dir=3 within, col=3.

Row8: col3. Next row7: col+dir=4 within, col=4.

Row7: col4. Next row6: col+dir=5 within, col=5.

Row6: col5 (rightmost). Next row5: attempt col+dir=6 out of bounds, so bounce: dir = -1, then col = col + dir = 5-1=4.

Row5: col4. Next row4: col+dir=3 within, col=3.

Row4: col3. Next row3: col+dir=2 within, col=2.

Row3: col2. Next row2: col+dir=1 wit

... [1129 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.60)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 875: "A wanted poster leads bounty hunters to fence in the fugitiv ... [2 chars truncated]" k=1 alts: 0/15 e2092e0c

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a 15×15 grid containing single-character color codes. In the top-left corner there is a distinctive cross or L-shaped structure made of X cells: the entire fourth row (row 3) is X X X X, and the entire fourth column (column 3) is X for rows 0 through 3. This creates a 4×4 region where row 3 and column 3 are all X, leaving a 3×3 sub-grid of non-X colored cells in rows 0–2, columns 0–2. This 3×3 sub-grid is the 'key pattern.' Somewhere else in the grid, a copy of this exact 3×3 pattern of colors exists at another location.

Do: 1. Extract the 3×3 key pattern from the top-left corner of the grid (rows 0–2, columns 0–2 — the non-X portion of the cross structure). 2. Search the rest of the grid for a 3×3 region that exactly matches this key pattern, cell by cell. 3. Once found, draw a 5×5 rectangular border of X cells centered on that 3×3 region. That is, if the matching 3×3 occupies rows r to r+2 and columns c to c+2, replace the border cells of the 5×5 rectangle spanning rows (r−1) to (r+3) and columns (c−1) to (c+3) with X. Specifically, the top and bottom rows of the 5×5 rectangle become all X, and the leftmost and rightmost columns of the three middle rows become X. The interior 3×3 cells (the matched pattern) remain unchanged. All other cells in the grid remain unchanged.

Grid: The grid is 15×15. It uses the standard ARC color palette (0–9) rendered as single characters: .=black, B=blue, R=red, G=green, Y=yellow, X=grey, M=magenta, O=orange, A=azure, W=maroon. The grid has no uniform background color; cells contain a variety of colors. The top-left 4×4 region always contains the cross template (row 3 all X, column 3 all X for rows 0–3). The output grid has the same dimensions as the input; the only change is the addition of the 5×5 X-border around the located pattern.

Figurative Descriptions (16 variants)

"A wanted poster leads bounty hunters to fence in the fugitive."

120b: k=1

See: In the top-left corner of the grid, a grey stone wall forms an L-shaped jail—bars running across the fourth row and down the fourth column. Locked behind those bars sits a 3×3 'mugshot' of colored suspects. Somewhere out in the wider grid, those same three-by-three suspects are hiding in plain sight among the crowd, wearing identical colors in the same arrangement.

Do: The sheriff studies the mugshot pinned to the jailhouse wall, then rides out into the grid to hunt for the fugitives. When the matching trio is found—same faces, same positions—the posse erects a grey stone fence around them: a 5×5 perimeter of X-bars that encircles the hideout one cell out in every direction. The fugitives inside remain untouched, but the fence announces they've been caught. Nothing else in the town changes—only the fence is new.

Grid: The town is a 15×15square of colored buildings. Grey (X) is the material of walls, bars, and fences. The jailhouse always occupies the top-left four rows and four columns, with its signature cross of grey bars. Every other color is a civilian going about their day. After the capture, the grid looks the same except for the new rectangular grey fence surrounding the discovered hideout.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (9952 chars)

We need to infer rule from example. Input and output differ only in certain area: In example, there is a 5x5 grey X fence added around a matching 3x3 pattern that appears both in jail (top-left 4x4 cross) and elsewhere.

Let's examine example 1. Input grid 15x15. The jail is top-left 4x4 with grey bars: rows 0-3, cols 0-3? Actually description: "grey stone wall forms an L-shaped jail—bars running across the fourth row and down the fourth column." So row index 3 (fourth row) all X across columns 0-3? And column index 3 (fourth column) all X down rows 0-3. In example input, we see first rows:

Row0: G Y . X . . G . X A . O . . .
Indices: col3 is X, col8 is X? Wait there is X at col3 and col8. But jail only first 4 columns? Actually row0 col3 X matches vertical bar. Row1 col3 X also. Row2 col3 X. Row3 row3 is X X X X (four X across). Indeed row3 (fourth row) columns 0-3 are X. So jail is 4x4 L shape: column 3 rows 0-3 X, row 3 columns 0-3 X.

Now inside the top-left 3x3 (rows0-2, cols0-2) there is a 3x3 pattern of colors (the mugshot). Let's extract from example input:

Rows0-2, Col0-2:
Row0: G Y .
Row1: . . Y
Row2: . A G

That's the 3x3 pattern.

Now elsewhere in grid there is same 3x3 pattern somewhere else. Need to locate. In output, they added a 5x5 X fence around that location (one cell border). Let's find where X's added.

Compare input vs output differences. Look at output rows 8-12 maybe.

In output row8 (index8): "B . . . B X X X X X . O . . G"
In input row8: "B . . . B . . . . . O . . . G"
So at row8, columns 5-9 become X's (5 X's). That's part of fence.

Row9 input: ". . B R . W G Y . . B . . R W"
Output row9: ". . B R . X G Y . X B . . R W"
So X at col5 and col9.

Row10 input: ". W . . A . . . Y . . M . A Y"
Output row10: ". W . . A X . . Y X . M . A Y"
X at col5 and col9.

Row11 input: "O O M . . . . A G . . . A R O"
Output row11: "O O M . . X . A G X . . A R O"
X at col5 and col9.

Row12 input: ". W . . R . . . . . . . B M"
Output row12: ". W . . R X X X X X . . . B M"
X at col5-9.

Thus fence is a 5x5 block of X covering rows8-12, cols5-9. Inside that block, the original 3x3 pattern should be at rows9-11, cols6-8? Let's check those cells:

Rows9-11, cols6-8 (center 3x3):
Row9 col6-8: G Y . (from output) Actually row9: "... X G Y . X ..." Wait row9 after fence: columns: let's index: col0 B,1 .,2 .,3 B? Wait row9 is ". . B R . X G Y . X B . . R W". Let's map indices: 0 .,1 .,2 B,3 R,4 .,5 X,6 G,7 Y,8 .,9 X,10 B,11 .,12 .,13 R,14 W. So center 3x3 (rows9-11, cols6-8) = rows9: G Y . ; row10: . . Y? Wait row10: ". W . . A X . . Y X . M . A Y". Indices: 0 .,1 W,2 .,3 .,4 A,5 X,6 .,7 .,8 Y,9 X,10 .,11 M,12 .,13 A,14 Y. So cols6-8: col6 ., col7 ., col8 Y. Row11: "O O M . . X . A G X . . A R O": indices: 0 O,1 O,2 M,3 .,4 .,5 X,6 .,7 A,8 G,9 X,10 .,11 .,12 A,13 R,14 O. So cols6-8: ., A, G.

Thus the 3x3 inside fence is:
Row9: G Y .
Row10: . . Y
Row11: . A G

Which matches the mugshot pattern exactly (G Y ., . . Y, . A G). Yes.

Thus rule:

... [6958 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 893: "The hottest ember escapes the dying campfires" k=1 alts: 7/15 e50d258f

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains two or three rectangular blocks of non-background cells, separated by borders of black (.) cells. Each block is composed of azure (A) cells forming a background within the block, with a small number of blue (B) and red (R) cells scattered among them.

Do: 1. Identify all distinct rectangular blocks of non-dot cells in the input grid. Blocks are separated from each other by rows or columns of dot cells. 2. For each block, count the number of red (R) cells it contains. 3. Select the block with the highest red cell count. 4. Extract that block exactly as it appears and output it.

Grid: The input grid is approximately 10x10 with black (.) cells serving as background and separators between blocks. Each block is a contiguous rectangle of non-dot cells. The output grid has the exact dimensions of the selected block, containing only the cells from that block with no surrounding dots.

Figurative Descriptions (16 variants)

"The hottest ember escapes the dying campfires"

120b: k=1

See: Scattered across the darkness like campfires in a valley, you see several glowing hearths — each a patch of pale azure coals with sparks of red flame and blue smoke drifting among them. The darkness between them is cold night air.

Do: Each campfire burns with a different intensity. Count the red-hot embers in each hearth — they are the true measure of its heat. The fire burning hottest, the one with the most living flame, refuses to die. Rescue it from the night and carry it forward whole.

Grid: The valley floor is dark (black). Each campfire is a compact rectangular bed of azure coals speckled with red embers and blue cinders. You must lift the hottest fire out intact — every coal, ember, and cinder in its exact arrangement — leaving the cold night behind.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10482 chars)

We need to infer rule from example. Input grid 10x10 with many colors. Output is 3x3 subgrid: A R R / R R B / R B A. So they extracted the "hottest fire": rectangular region (campfire) with azure (A) coals, red (R) embers, blue (B) cinders. Need to find region with most red count? Let's analyze example input to locate such region.

Let's map coordinates. I'll write rows with indices 0-9.

Row0: B B B A . . . . . .
Row1: B A B B . B A A B A
Row2: A R A B . A B A R A
Row3: B B B A . A A A A B
Row4: A B A A . A B R A R
Row5: . . . . . A A A B A
Row6: . . . . . B B A B A
Row7: . A R R . A B B A R
Row8: . R R B . . . . . .
Row9: . R B A . . . . . .

We need to find rectangular clusters of A,R,B maybe contiguous? Let's locate region that matches output 3x3: pattern:
A R R
R R B
R B A

Search in input. Look at rows7-9 columns? Row7: . A R R . A B B A R
Row8: . R R B . . . . . .
Row9: . R B A . . . . . .

If we take columns1-3? Row7 col1=A, col2=R, col3=R => matches first row A R R. Row8 col1=R, col2=R, col3=B => matches second row R R B. Row9 col1=R, col2=B, col3=A => matches third row R B A. Yes! So the 3x3 region at rows7-9, cols1-3 is output.

Thus rule: find rectangular region (maybe minimal bounding rectangle) that contains only colors A,R,B (azure, red, blue) and maybe is the "hottest" meaning most reds count. Among all such rectangular "campfires" (compact rectangular bed of azure coals speckled with red embers and blue cinders). Need to extract the one with highest number of red cells. If tie maybe something else.

Let's verify there are other A,R,B rectangles elsewhere. For example rows0-3 columns0-3 have many B and A but also maybe other colors? Row0 col0-3: B B B A (contains B and A, no R). Row1 col0-3: B A B B (no R). Row2 col0-3: A R A B (has R). Row3 col0-3: B B B A (no R). That's a 4x4 region but includes rows with no R maybe still qualifies? But maybe campfire must be contiguous region of only A,R,B (no other colors). The region we found rows7-9 cols1-3 includes only A,R,B (no other colors). Check other possible rectangles: rows2-4 cols0-3? Row2: A R A B (A,R,B). Row3: B B B A (B,A). Row4: A B A A (A,B). Contains only A,R,B. That's 3x4 region (rows2-4, cols0-3). Count reds: row2 has 1 R, row3 none, row4 none => 1 red. The selected region has reds count: rows7-9: row7 has 2 R, row8 has 2 R, row9 has 1 R => total 5 reds. So indeed highest.

Thus algorithm: Find all maximal rectangular subgrids that consist solely of the three colors A,R,B (no other colors). Among them, select the one with the greatest number of R cells. Output that subgrid exactly.

But need to consider "compact rectangular bed" maybe means any rectangle (not necessarily maximal) that contains only those colors. Could be many overlapping. The one with most reds likely is the rectangle that encloses the region with most reds, maybe the minimal bounding rectangle of the connected component of A,R,B? In example, the region rows7-9 cols1-3 is a connected compone

... [7542 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 898: "A photograph printed with extra-thick borders, like a Polaro ... [28 chars truncated]" k=2 alts: 3/15 e633a9e5

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: A 3x3 grid of colored cells. Each cell contains one of several possible colors. The grid has three rows and three columns, with no special background color or distinguished elements — every cell is equally significant.

Do: Expand the 3x3 input grid into a 5x5 output grid by duplicating the border rows and columns while keeping the center row and column single. Specifically: (1) Duplicate each column: the first column appears twice, the middle column appears once, and the last column appears twice, producing 5 columns per row. (2) Duplicate each row: the first row appears twice, the middle row appears once, and the last row appears twice, producing 5 rows total. In index terms, output row i maps to input row [0,0,1,2,2][i], and output column j maps to input column [0,0,1,2,2][j].

Grid: Input is always a 3x3 grid. Output is always a 5x5 grid. There is no background color; all cells carry meaningful color values. The transformation preserves all original colors and spatial relationships, only scaling the grid by stretching the border rows and columns.

Figurative Descriptions (16 variants)

"A photograph printed with extra-thick borders, like a Polaroid that bleeds at the edges."

120b: k=2

See: A tiny mosaic tile, three paces by three paces — a miniature world where every colored stone carries equal weight, none more important than another.

Do: The edges of the mosaic are sticky — they cling to the frame and leave an echo of themselves as the tile is stretched onto a larger wall. The outermost stones on every side press themselves into the plaster twice, as if they're bracing against the border, while the center stones, confident in their position, sit calmly without doubling. Think of bread dough rolled outward: the crust thickens at the rim while the middle stays thin.

Grid: The miniature mosaic is always three-by-three. The wall it's mounted on is always five-by-five. No stone changes color during the stretching — they simply leave shadows of themselves along the edges, like footprints pressed deeper at the margins of a path.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (7407 chars)

We need to infer rule: Input 3x3, output 5x5. The description: edges double (appear twice) while center stays single. Essentially each cell of input is expanded to a 2x2 block if it's on outer border? Let's examine example1.

Input:
Row1: B G X
Row2: B R A
Row3: A G A

Output 5x5:
Row1: B B G X X
Row2: B B G X X
Row3: B B R A A
Row4: A A G A A
Row5: A A G A A

Let's map positions. Output coordinates (0-4). Input coordinates (0-2). The rule maybe: For each input cell, we place it in output at positions (i*? + offset). Since output larger by 2 rows/cols. Possibly each input cell maps to a 2x2 block, but interior cells (center) maybe map to 1x1? Let's test.

Input (0,0) B (top-left corner). In output, top-left 2x2 block (rows0-1, cols0-1) are B B; B B. Indeed output rows0-1 cols0-1 are B. Good.

Input (0,1) G (top edge, not corner). Should expand to 2x? maybe 2 rows, 1 column? Let's see output rows0-1, col2? Output col2 is G at rows0-1? Row0 col2 = G, row1 col2 = G. Yes column2 (index2) has G for rows0-1. So a 2x1 vertical block.

Input (0,2) X (top-right corner). Output rows0-1, cols3-4? Row0 col3 X, col4 X; row1 col3 X, col4 X. Yes 2x2 block.

Input (1,0) B (left edge middle). Output rows2-? Actually row index 2? Let's see output row2 col0-1? Row2 col0 B, col1 B. Row3 col0 A? Wait row3 col0 A not B. So B appears only rows2? Actually output row2 col0 B, col1 B. Row3 col0 A. So B occupies a 1x2 horizontal block? Let's examine: Input (1,0) is left edge middle. Output rows2 (index2) col0-1 are B B. Row3 col0-1 are A A (different). So B appears only in row2, not row3. So left edge middle expands horizontally (1 row, 2 columns). Similarly input (1,2) A (right edge middle) appears in output row2 col3-4? Row2 col3 A, col4 A. Yes.

Input (2,0) A (bottom-left corner) appears rows3-4 cols0-1 as A A; A A. Yes 2x2.

Input (2,1) G (bottom edge middle) appears rows3-4 col2? Row3 col2 G, row4 col2 G. That's 2x1 vertical.

Input (2,2) A (bottom-right corner) rows3-4 col3-4 A A; A A.

Now interior cell (1,1) R (center). Output row2 col2 = R? Row2 col2 is R. Only single cell. Yes.

Also interior cell (1,2) A? Wait (1,2) is A (right middle edge) not interior. interior only (1,1). Also (1,0) left edge middle.

Thus rule: Expand 3x3 to 5x5 by duplicating border cells outward: corners become 2x2 blocks, edge (non-corner) cells become 2x1 or 1x2 blocks depending orientation, and center stays 1x1.

Specifically mapping: Output size = input size + 2 (since each side adds one extra column/row). For each input cell (i,j):
- If i==0 (top row) and j==0 (left col) => corner: occupies rows 0-1, cols 0-1.
- If i==0 and j==2 (right corner) => rows0-1, cols3-4.
- If i==2 and j==0 => rows3-4, cols0-1.
- If i==2 and j==2 => rows3-4, cols3-4.
- If i==0 (top edge, not corner) => rows0-1, col j+1? Actually j=1 maps to col2 (single column). So top edge cells become vertical 2x1 block at column j+1 (since corners occupy 0-1 and 3-4). So column index = j+1.
- If i==2 (bottom

... [4503 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.04)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 910: "A dance troupe passes choreography down the line at each reh ... [59 chars truncated]" k=1 alts: 3/15 e78887d1

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid is organized into two or more horizontal groups of 3×3 colored blocks, separated by rows of all-black cells. Within each group, there are N colored blocks (e.g., R, G, B, X) arranged left to right, each occupying a 3×3 region and separated by a single column of black cells. Every block uses exactly one non-black color but has its own unique pattern of filled and empty cells within its 3×3 space. The same set of colors appears in the same left-to-right order in every group. From one group to the next, the patterns (filled/empty arrangements) shift cyclically among the color positions: each color adopts the pattern that belonged to the color immediately to its right in the previous group, with the rightmost color's pattern wrapping around to the leftmost color.

Do: 1. Identify the horizontal groups separated by all-black rows. Each group is exactly 3 rows tall. 2. Within each group, identify the N colored 3×3 blocks arranged left to right, noting their color order (e.g., R, G, B, X). 3. Extract the 3×3 pattern (which cells are filled vs. black) for each color block in the last group. 4. Produce the output by applying one more cyclic left-shift of patterns: assign to each color position the pattern that belonged to the color one step to its right in the last group. The rightmost color receives the pattern from the leftmost color. 5. Render the output as a single 3-row group with the same color order and column spacing, filling each color's 3×3 region with its new pattern in that color.

Grid: The input grid has a width of (N × 3 + (N−1)) columns, where N is the number of colored blocks per group, and a height of (G × 3 + (G+1)) rows, where G is the number of groups (the extra rows are all-black separator rows including top and bottom borders). The output grid is always exactly 3 rows tall and the same width as the input, containing a single group with no surrounding black separator rows. The background color is black (.).

Figurative Descriptions (16 variants)

"A dance troupe passes choreography down the line at each rehearsal, and the final performance reveals the next routine."

120b: k=1

See: Troupes of colored dancers—R, G, B, and sometimes X—stand in a line on a dark stage, each performing their own signature 3×3 choreography. The stage is divided into rehearsal rounds stacked vertically, separated by blackout curtains. In each new round, every dancer has learned the moves of the neighbor to their right, while the dancer on the far right loops back to mimic the one on the far left—a rolling relay of stolen choreography passing through the lineup.

Do: Watch the dancers rehearse through all their rounds, paying close attention to the final rehearsal. Now predict the opening night performance: each dancer steps forward wearing their own color but performing the choreography that belonged to the dancer one step to their right in that last rehearsal. The dancer at the end of the line inherits the first dancer's routine, completing the circle. The curtain rises on a single, final performance—no blackouts, no intermissions.

Grid: The dark stage spans enough room for each dancer's 3×3 spotlight with a single dark column between them. Blackout rows frame each rehearsal round top and bottom. Opening night strips away all the darkness—just one three-row performance under the lights, same stage width, no curtains.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (13721 chars)

We need to understand the puzzle. Input grid includes multiple "rehearsal rounds" separated by rows of all '.' (black). In example 1, input has 5 rows? Let's examine.

Example 1 Input (as shown):
". . . . . . . . . . . . . . ."
"R R R . G . . . B . . . X . X"
". . R . G . . . . B B . X . X"
". . R . G G G . B . . . X . X"
". . . . . . . . . . . . . . ."

So 5 rows, 15 columns. The pattern: first row all '.' (top blackout). Then three rows of data (the three rows of each dancer's 3x3 choreography? Actually each dancer occupies a 3x3 block with a column of '.' between dancers). Then bottom row all '.' (bottom blackout). So there are 3 data rows (height of choreography). The output is 3 rows (no black rows). Output given:

"R . . . G . . . B . B . X X X"
"R . . . . G G . B . B . . . X"
"R R R . G . . . B . B . . . X"

That's also 3 rows, 15 columns.

So transformation: For each dancer column (i.e., each 3x3 block), we need to shift the block's pattern to the left? Actually description: each dancer inherits choreography of dancer to right in final rehearsal. The final rehearsal is the last round before output? In example, there is only one round? Actually input has only one round (since only one set of data rows). But description says multiple rounds stacked vertically separated by blackout rows. In example 1, there is only one round (the three rows). So the final rehearsal is that same round. Then output: each dancer's 3x3 block is replaced by the block of the dancer to its right (wrapping). Let's test.

Identify dancer blocks: In input, we have columns with colors: Let's map columns indices 0-14.

Row1 (index1): "R R R . G . . . B . . . X . X"
Row2: ". . R . G . . . . B B . X . X"
Row3: ". . R . G G G . B . . . X . X"

We need to parse 3x3 blocks separated by a column of '.' (maybe multiple '.'?). Let's locate '.' columns that separate. At column 3 is '.' after first block? Let's see columns:

col0 R, col1 R, col2 R, col3 ., col4 G, col5 ., col6 ., col7 ., col8 B, col9 ., col10 ., col11 ., col12 X, col13 ., col14 X

Seems pattern: each dancer's block width 3 columns, then a single '.' column separator. So first dancer occupies cols0-2 (R,R,R; .,.; .,.)? Actually rows: first dancer block (3x3) should be:

Rows:
R R R
. . R
. . R

That's a shape: a vertical line on right side? That's the block for dancer 1 (color maybe R?). Indeed block contains R's only.

Second dancer block columns 4-6? Let's see col4 G, col5 ., col6 .; row2 col4 G, col5 ., col6 .; row3 col4 G, col5 G, col6 G. So block:

G . .
G . .
G G G

That's a shape for G dancer.

Third dancer block columns 8-10: col8 B, col9 ., col10 .; row2 col8 ., col9 B, col10 B; row3 col8 B, col9 ., col10 . . So block:

B . .
. B B
B . .

That's B dancer.

Fourth dancer block columns 12-14: col12 X, col13 ., col14 X; row2 col12 X, col13 ., col14 X; row3 col12 X, col13 ., col14 X. So block:

X . X
X . X
X . X

That's X dancer.

Now output rows: Let's extract blocks after transformation.

Output rows:

Row1

... [10790 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.02)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 916: "A night sky remembers each river it has swallowed, plus the ... [31 chars truncated]" k=1 alts: 1/15 e872b94a

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid has a black (.) background with grey (X) cells that form one or more snake-like connected components. Each component is a path one cell wide, meaning X cells connect to each other orthogonally (up, down, left, right) to form continuous, non-branching shapes. The components are spatially separated from one another — no X cell in one component is orthogonally adjacent to an X cell in another.

Do: 1. Identify all grey (X) cells in the input grid. 2. Determine the number of distinct connected components formed by these X cells, where two X cells are connected if they are orthogonally adjacent (share an edge horizontally or vertically). 3. Count the total number of connected components and add 1 to that count. 4. Produce an output grid that has exactly that many rows (count + 1), each row containing a single black (.) cell.

Grid: The output grid is always a single column (width 1). Its height equals the number of connected components of grey cells in the input plus one. Every cell in the output is black (.).

Figurative Descriptions (16 variants)

"A night sky remembers each river it has swallowed, plus the darkness that was there before."

120b: k=1

See: Against a vast dark sky, grey rivers wind across the landscape — each one a single, unbroken current that never forks or touches another. They twist and turn like serpents asleep on the void, each keeping a respectful distance from its neighbors.

Do: The sky is an ancient scribe who tallies every river it has absorbed. For each grey river swallowed into the darkness, the scribe places one drop of ink on a scroll. But the scribe always begins with one drop already on the page — a mark for the original emptiness that existed before any river flowed. Roll up the scroll when the counting is done.

Grid: The scroll is a single column of pure night. Its length is the scribe's tally — one drop for the void, one for each river the sky consumed.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (7866 chars)

We need to infer rule. Input grid contains '.' black and 'X' grey, maybe other colors not present. Output is a single column of '.'? In example 1, input is 7x7 grid with '.' and X. Output is three rows each '.'? Actually output shown:
.
.
.
So three rows, each a single '.' (black). That matches description: count of grey rivers (continuous lines) plus one for void. Need to count number of grey rivers (connected components) that are single unbroken lines that never fork or touch another. In example, we need to count number of such rivers. Output length = count + 1? Wait description: scribe always begins with one drop already on the page — a mark for the original emptiness that existed before any river flowed. So total drops = 1 + number of rivers. Output is a column of that many '.'? In example output length is 3, so there were 2 rivers in input.

Let's verify: Input 7x7 grid. Let's map X positions.

Row0: . X . . . . . (col1)
Row1: . X X . . . . (col1,2)
Row2: . . X . . X X (col2,5,6)
Row3: . X X . . X . (col1,2,5)
Row4: . X . . X X . (col1,4,5)
Row5: . X . . X . . (col1,4)
Row6: . X . . X . . (col1,4)

We need to identify "grey rivers": each is a single, unbroken current that never forks or touches another. Likely each river is a connected component of X cells that forms a path (maybe orthogonal adjacency). Let's find components.

Component A: column1 (col1) many X's rows0-6, plus some adjacent X at row1 col2, row2 col2, row3 col2? Actually row1 col2 is X adjacent to col1 row1 (horizontal). Row2 col2 X adjacent to row1 col2 (vertical). Row3 col2 X adjacent to row2 col2. So that forms a shape maybe branching? At row3 col2 also adjacent to row3 col1 (X). So component includes many cells: basically a thick shape maybe branching? But rule says rivers never fork or touch another. This component may have a branch? Let's see connectivity: Starting at (0,1) downwards column1. At row1, there is also (1,2) to the right, then (2,2) down, then (3,2) down, also (3,1) left. So shape like a T? At (3,2) also maybe (3,5) far away not connected. So component includes a branch at (1,2) and (2,2) and (3,2) making a line that goes right then down then left? That's still a single path without forks? Actually there is a branch at (1,2) which is a dead end? The path could be: start at top (0,1) down to (3,1) then right to (3,2) then up to (2,2) then up to (1,2) then maybe stop. That's a snake shape, no branching (each cell degree <=2). Let's check degrees: (0,1) degree1 (down). (1,1) degree2 (up/down). (2,1) degree2. (3,1) degree2 (up/right). (3,2) degree3? Adjacent to (3,1) left, (2,2) up, (4,2?) not X, (3,3?) none. So degree2 actually (left and up). Wait also (1,2) adjacent to (2,2) down, (1,1) left? Actually (1,2) adjacent left to (1,1) which is X? row1 col1 is X yes. So (1,2) degree2 (left and down). (2,2) degree2 (up/down). So all degrees <=2, so it's a simple path (maybe with bends). So that's one river.

Component B: there are X's at columns5-6 etc. 

... [4965 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.33)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 921: "Four royal houses compete for territory, and the crown goes ... [35 chars truncated]" k=3 alts: 7/15 e99362f0

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Example 6

Input

→

Output

Literal Description

See: The input is a rectangular grid divided into four quadrants by a yellow cross — a horizontal row of all yellow cells and a vertical column of all yellow cells that intersect at one point. This cross splits the grid into four equally-sized rectangular regions. Each quadrant contains a pattern made from one signature color on a black background: the top-left quadrant uses orange (O), the top-right uses white (W), the bottom-left uses red (R), and the bottom-right uses azure (A). Each quadrant has the same dimensions.

Do: 1. Identify the yellow cross (one full yellow row and one full yellow column) and extract the four quadrants, ignoring the yellow dividers. All four quadrants share the same dimensions — this will be the output size. 2. For each cell position in the output grid, examine the corresponding cell in all four quadrants. Each quadrant cell is either its signature color or black. 3. Determine which quadrants have a non-black (colored) value at that position. If none do, the output cell is black. 4. If one or more quadrants are non-black at that position, select the winner using the fixed priority order: azure (A) > orange (O) > white (W) > red (R). The highest-priority non-black color wins and is placed in the output cell.

Grid: The input grid is 11 rows by 9 columns, with a yellow row (row 5) and a yellow column (column 4) forming a cross that divides the grid into four 5×4 quadrants. The output grid matches the quadrant dimensions: 5 rows by 4 columns. The background color is black (.). The four signature colors — orange, white, red, azure — each appear exclusively in their respective quadrant in the input but may all appear in the output.

Figurative Descriptions (16 variants)

"Four royal houses compete for territory, and the crown goes to whoever has the strongest claim."

120b: k=3

See: A golden cross — like two roads meeting at a crossroads — carves the land into four kingdoms. Each kingdom flies a single banner over its domain: orange in the northwest, white in the northeast, red in the southwest, and azure in the southeast. Within each kingdom, some plots are settled (colored) and some lie fallow (black).

Do: The golden borders dissolve, and all four kingdoms collapse onto the same contested map. Where no kingdom has planted a flag, the land remains barren. Where multiple houses claim the same plot, the dispute is settled by ancient right of succession: the Azure throne outranks all, Orange bows only to Azure, White yields to both Orange and Azure, and Red — the least of the four — wins only when no other house stands on the ground.

Grid: The golden crossroads splits the realm into four provinces of equal size. When the borders fall, the contested map matches the size of a single province. Every plot on this final map bears the color of whichever surviving house holds the strongest claim there.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (7839 chars)

We need to understand puzzle. The description: golden cross splits grid into four quadrants (kingdoms). Input grid includes a central row/column of Y (yellow) representing the golden cross? In examples, there is a row of Y across middle and column of Y across middle. Indeed each input shows a 11x9? Let's examine Example1 input dimensions: lines? Let's count rows: appears 11 rows? Actually list: 
Row1 ". O . . Y W . W ."
Row2 "O O . . Y W . . ."
Row3 ". . . . Y W . W W"
Row4 ". O O O Y . . . ."
Row5 ". . O O Y . . W W"
Row6 "Y Y Y Y Y Y Y Y Y"
Row7 ". R R . Y A A . ."
Row8 "R R . R Y A . A A"
Row9 "R . R R Y . A . A"
Row10 "R . R R Y . A A ."
Row11 "R . . . Y . . A ."

So 11 rows, 9 columns. The central row (row6) all Y. Also central column (col5) all Y. So cross splits into four quadrants each of size 5 rows x 4 columns? Actually above cross, there are 5 rows (rows1-5) and below cross 5 rows (rows7-11). Left side columns1-4, right side columns6-9. So each quadrant 5x4.

Goal: "golden borders dissolve, and all four kingdoms collapse onto the same contested map. The contested map matches the size of a single province." So output grid size is 5 rows x 4 columns (size of one quadrant). Indeed example outputs have 5 rows and 4 columns.

Now rule: For each cell position within a quadrant (relative coordinates), we look at the four quadrants' cells at that position (same offset from cross). Each quadrant has its own color (orange, white, red, azure) as "banner" meaning the dominant house for that quadrant? Actually description: each kingdom flies a single banner over its domain: orange in the northwest, white in the northeast, red in the southwest, azure in the southeast. Within each kingdom, some plots are settled (colored) and some lie fallow (black). So the colors present in each quadrant are the "settled" cells (non-black). The banner color maybe not directly present; but the rule for conflict: Azure outranks all, Orange bows only to Azure, White yields to both Orange and Azure, Red wins only when no other house stands.

Thus for each position, we consider the colors present in each quadrant at that relative position. If multiple non-black colors, we select according to hierarchy: Azure > Orange > White > Red. If none (all black), output black (.)? The description says "Where no kingdom has planted a flag, the land remains barren." So black.

Thus we need to map colors: Azure = A, Orange = O, White = W, Red = R. Also there may be other colors? In examples only these four appear plus Y (cross) and . (black). So rule: For each cell (i,j) in output (size 5x4), look at four quadrants:

NW quadrant: rows 1-5, cols1-4 (relative positions (r,c))
NE: rows1-5, cols6-9
SW: rows7-11, cols1-4
SE: rows7-11, cols6-9

Take the character at each quadrant cell. If it's '.' (black) treat as empty. Among non-'.' choose highest rank: A > O > W > R. Output that character. If multiple same highest? they are same char anyway.

Let's test with Example1 to

... [4983 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.09)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 931: "Four voices singing in a round, but the strongest voice alwa ... [17 chars truncated]" k=1 alts: 6/15 ea9794b1

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Example 6

Input

→

Output

Literal Description

See: The input is a 10x10 grid divided into four 5x5 quadrants. Each quadrant uses exactly one non-black color paired with black (.) cells: Top-Left uses Yellow (Y), Top-Right uses Green (G), Bottom-Left uses Maroon (W), and Bottom-Right uses Azure (A). Each quadrant forms a sparse pattern of its color against a black background.

Do: To produce the 5x5 output grid, overlay all four quadrants cell by cell. For each position (r, c) where r ranges 0-4 and c ranges 0-4, examine the corresponding cell in each quadrant: Top-Left at (r, c), Top-Right at (r, c+5), Bottom-Left at (r+5, c), Bottom-Right at (r+5, c+5). Each quadrant cell is either its designated color or black. Assign the output cell using this strict priority order from highest to lowest: Green (G) > Maroon (W) > Azure (A) > Yellow (Y) > black (.). That is, if Green is present in any quadrant at that position, output Green. Otherwise if Maroon is present, output Maroon. Otherwise if Azure is present, output Azure. Otherwise if Yellow is present, output Yellow. Otherwise output black (.).

Grid: The input grid is always 10x10, composed of four non-overlapping 5x5 quadrants arranged in a 2x2 layout. The output grid is always 5x5. Each input quadrant contains exactly one chromatic color (Yellow, Green, Maroon, or Azure) plus black cells. The output uses all four colors plus black.

Figurative Descriptions (16 variants)

"Four voices singing in a round, but the strongest voice always wins the note."

120b: k=1

See: A stage is divided into four equal chambers. In each chamber, a single singer performs against silence — Yellow hums in the top-left, Green belts from the top-right, Maroon growls from the bottom-left, and Azure whispers from the bottom-right. Each performer's song is scattered across their chamber: some notes ring out, others are rests.

Do: The four chambers collapse into one shared stage, and at every position the singers compete for who is heard. Green has the most powerful voice and drowns out everyone. Maroon cuts through all but Green. Azure overpowers only the timid Yellow. Yellow is barely audible, singing only where no one else stands. Where all four singers rest, the audience hears nothing but silence.

Grid: The full concert hall is a 10×10 space with four 5×5 sound booths arranged in a square. When the walls come down, a single 5×5 performance remains — a chord at each position determined by whichever voice carries the farthest. The final song shimmers with all four colors and pockets of silence.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (8804 chars)

We need to infer rule. Input is 10x10 with four 5x5 quadrants: top-left (rows0-4, cols0-4), top-right (rows0-4, cols5-9), bottom-left (rows5-9, cols0-4), bottom-right (rows5-9, cols5-9). Output is 5x5 representing merged quadrant according to priority: Green > Maroon > Azure > Yellow > empty (.)? Let's verify with example.

We need to map colors: G=green, W=maroon, A=azure, Y=yellow, . black (silence). In description: Green most powerful, drowns out everyone. Maroon cuts through all but Green. Azure overpowers only Yellow. Yellow only where no one else stands. Where all four rest, silence.

Thus priority order: Green > Maroon > Azure > Yellow > . (silence). Also if multiple same? Not needed.

Now we need to combine corresponding cells from four quadrants: For each position (i,j) within 5x5, look at four cells: TL[i][j], TR[i][j], BL[i][j], BR[i][j]. Determine which color appears according to priority. If a cell is '.' (black) meaning no note. So we need to pick highest priority non-'.' among the four. If all '.' then '.'.

Let's test with example 1 to ensure matches output.

We need to extract quadrants from example input.

Example Input 10x10 lines:

Row0: ". Y Y Y . . . . G G"
Row1: "Y Y . . . G . G G ."
Row2: "Y . . Y Y . G G G ."
Row3: ". . Y . Y G . . G ."
Row4: ". . Y Y Y G G G G G"
Row5: ". W . W W . . . A ."
Row6: "W . . W W . A A . A"
Row7: ". . . W . . . A A ."
Row8: ". . W W . A . A . ."
Row9: "W W . W . . A . A A"

Let's index rows 0-9, cols0-9.

Now TL quadrant rows0-4 cols0-4:

Rows0-4, cols0-4:
r0: . Y Y Y .
r1: Y Y . . .
r2: Y . . Y Y
r3: . . Y . Y
r4: . . Y Y Y

TR quadrant rows0-4 cols5-9:
r0 cols5-9: . . . G G
r1: G . G G .
r2: . G G G .
r3: G . . G .
r4: G G G G G

BL quadrant rows5-9 cols0-4:
r5: . W . W W
r6: W . . W W
r7: . . . W .
r8: . . W W .
r9: W W . W .

BR quadrant rows5-9 cols5-9:
r5: . . . A .
r6: . A A . A
r7: . . A A .
r8: A . A . .
r9: . A . A A

Now compute output 5x5 using priority.

Let's list each position (i,j) 0-4.

I'll create table.

i0 row0:

j0: TL '.' ; TR '.' ; BL '.' ; BR '.'? Wait BR row5 col5? Actually position (0,0) corresponds to rows0 TL, rows0 TR, rows5 BL, rows5 BR? No mapping: For each i (0-4) corresponds to row i in TL and row i in TR; for BL and BR rows i+5. Similarly columns j.

So for (0,0):
TL r0c0 = '.'
TR r0c5 = '.' (col5)
BL r5c0 = '.' (row5 col0)
BR r5c5 = '.' (row5 col5)
All '.' => output '.' matches example output row0 col0 '.' yes.

j1 (0,1):
TL r0c1 = 'Y'
TR r0c6 = '.' (col6)
BL r5c1 = 'W'
BR r5c6 = '.' (col6)
Now priority: Green > Maroon (W) > Azure (A) > Yellow (Y). No G, have W and Y. Maroon higher than Yellow, so output should be W. Example output row0 col1 = 'W' yes.

j2 (0,2):
TL r0c2 = 'Y'
TR r0c7 = '.' (col7)
BL r5c2 = '.' (row5 col2)
BR r5c7 = '.' (row5 col7)
Only Y => output Y. Example row0 col2 = 'Y' yes.

j3 (0,3):
TL r0c3 = 'Y'
TR r0c8 = 'G'
BL r5c3 = 'W'
BR r5c8 = 'A'
Now we have G, W, A, Y. Highest priority G => output G. 

... [5990 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.24)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 939: "A secret society's signet ring stamps its color onto a wax s ... [4 chars truncated]" k=3 alts: 14/15 ed74f2f2

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Example 6

Input

→

Output

Literal Description

See: The input is a 5×9 grid with a black background. Embedded within it are two 3×3 patterns made of grey (X) cells, separated by a column of black cells. The left pattern occupies rows 1-3, columns 1-3, and the right pattern occupies rows 1-3, columns 5-7. Each pattern is a distinct arrangement of grey cells against the black background.

Do: 1. Extract the left 3×3 subgrid (rows 1-3, columns 1-3) and the right 3×3 subgrid (rows 1-3, columns 5-7). 2. Identify the shape of the left subgrid to determine the output color: (a) If the left pattern is a T-shape (top row all filled, middle and bottom rows only center filled: XXX / .X. / .X.), the color is Blue (B). (b) If the left pattern is an S-shape (top row right two filled, middle row center filled, bottom row left two filled: .XX / .X. / XX.), the color is Green (G). (c) If the left pattern is a Z-shape (top row left two filled, middle row center filled, bottom row right two filled: XX. / .X. / .XX), the color is Red (R). 3. Create a 3×3 output grid by taking the right subgrid and replacing every grey (X) cell with the determined color. Black (.) cells remain black.

Grid: The input grid is always 5 rows by 9 columns with a black (.) background. Row 0 and row 4 are entirely black, forming borders. The two 3×3 patterns are always located at fixed positions: left pattern at rows 1-3, columns 1-3 and right pattern at rows 1-3, columns 5-7, separated by a black column (column 4). The output grid is always 3×3.

Figurative Descriptions (16 variants)

"A secret society's signet ring stamps its color onto a wax seal."

120b: k=3

See: A dark hall holds two artifacts side by side on a velvet display, separated by a thin shadow. On the left sits a signet ring — a grey emblem whose silhouette belongs to one of three ancient orders. On the right lies a blank wax seal, also shaped in grey, waiting to receive the ring's impression.

Do: Read the heraldry of the left signet ring to learn which order it belongs to — each order claims a sacred color. Then press the ring into the wax seal on the right: wherever the seal bears grey, it absorbs the order's color, and wherever the seal is empty darkness, it stays untouched. The seal, now branded, is all that remains.

Grid: The display case is a narrow corridor of darkness. The two artifacts always sit in the same positions within it, like twin pedestals in a museum wing, divided by a single strip of shadow. Only the branded seal — small and square — is carried out of the hall.

gpt-oss-120b — MARC valid, solved with 3 examples

Input

→

Prediction

vs

Expected

Reasoning (5937 chars)

We need to infer rule. Input grids are 5 rows x 9 columns? Actually examples show 5 rows, 9 columns. Output grids are 3x3? Yes output is 3 rows, 3 columns. So transformation extracts a 3x3 region (the "seal") on right side, and fills with colors based on left artifact (ring) order color.

We need to map: left artifact shape (grey X pattern) determines a color (order's sacred color). Then the right artifact shape (grey X pattern) is the mask: wherever there is X (grey) in right artifact, output gets that color; wherever there is . (black) stays black (represented as .). Output is 3x3.

Thus we need to identify left and right artifact positions. The corridor is narrow, with two artifacts side by side separated by a single column of shadow (maybe a column of .?). In examples, the left artifact appears in columns 2-4? Let's examine Example1 input.

I'll write coordinates. Row1 all . (top border). Row2: ". X X X . X . X ." Let's index columns 0-8.

Row2: col0 '.'; col1 'X'; col2 'X'; col3 'X'; col4 '.'; col5 'X'; col6 '.'; col7 'X'; col8 '.'

Row3: ". . X . . X X X ." => col0 '.';1 '.';2 'X';3 '.';4 '.';5 'X';6 'X';7 'X';8 '.'

Row4: ". . X . . X X . ." => col0 '.';1 '.';2 'X';3 '.';4 '.';5 'X';6 'X';7 '.';8 '.'

So left artifact likely columns 1-3? Let's see pattern of left artifact shape: In example1 left side (columns 1-3) rows 2-4:

Rows2-4, cols1-3:
Row2: X X X
Row3: . X .
Row4: . X .

That's a shape: a vertical line of X at col2 rows2-4, plus top row full X across 1-3. That's like a T shape? Actually looks like a "signet ring" shape maybe representing order.

Right artifact likely columns 5-7? Let's extract columns5-7 rows2-4:

Row2 col5 X, col6 ., col7 X => X . X
Row3 col5 X, col6 X, col7 X => X X X
Row4 col5 X, col6 X, col7 . => X X .

That's shape: top row X . X, middle row X X X, bottom row X X . That's like a different shape.

Now output for example1 is:
B . B
B B B
B B .

Corresponds to right artifact shape but with color B (blue) where X present, and . where . present. Indeed compare: right artifact pattern (X . X / X X X / X X .) matches output B . B / B B B / B B . So rule: output = right artifact pattern colored with order's color.

Now need to determine order's color based on left artifact shape. In example1 left shape (T shape) maps to color B (blue). Example2 left shape? Let's extract left shape for example2.

Example2 input rows2-4 columns? Row2: ". . X X . X . X ." => columns: 0 '.';1 '.';2 'X';3 'X';4 '.';5 'X';6 '.';7 'X';8 '.'

Row3: ". . X . . X . X ." => col2 X, col3 ., col4 ., col5 X, col6 ., col7 X

Row4: ". X X . . X X . ." => col1 X, col2 X, col3 ., col4 ., col5 X, col6 X, col7 ., col8 .

Left artifact columns maybe 1-3? Let's see columns1-3 rows2-4:

Row2 col1 '.'; col2 'X'; col3 'X' => . X X
Row3 col1 '.'; col2 'X'; col3 '.' => . X .
Row4 col1 'X'; col2 'X'; col3 '.' => X X .

Pattern: . X X / . X . / X X . That's a shape maybe different order.

Output is G . G / G . G / G G . (from exam

... [3003 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 3 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 941: "A talent scout watches a parade from a balcony, then reports ... [32 chars truncated]" k=1 alts: 8/15 edcc2ff0

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The grid is divided into two regions by a solid horizontal row of a single color at row 7. Above that divider (rows 0–6) is a mostly black section with exactly three single-pixel colored indicators placed in column 0 at rows 1, 3, and 5. Below (rows 7–19) is a rectangular grid whose background color matches the divider row; scattered across this lower grid are small connected shapes in various colors, some matching the indicator colors and some not.

Do: 1. Identify the three indicator colors from column 0 at rows 1, 3, and 5 in the top section. These are the 'kept' colors. 2. Identify the background color of the lower grid (the color filling the solid divider at row 7). 3. In the lower grid (rows 7–19), find all connected components of each kept color (4-connected adjacency). Count how many distinct components exist for each kept color. 4. Update the top section: for each indicator at row r, extend that color rightward from column 0 so the total number of filled cells in that row equals the component count for that color. If the count is 0, clear the indicator entirely (replace with black). All other cells in the top section remain black. 5. In the lower grid, remove every shape whose color is NOT one of the three kept colors by replacing those cells with the background color. Shapes in kept colors remain unchanged.

Grid: The grid is always 10 columns wide and 20 rows tall. Row 7 is a solid horizontal divider whose color defines the background of the lower 13-row region (rows 7–19). The upper 7 rows (0–6) use black as background, with indicator pixels only at (row 1, col 0), (row 3, col 0), and (row 5, col 0). Odd rows in the top section hold indicators; even rows are entirely black.

Figurative Descriptions (16 variants)

"A talent scout watches a parade from a balcony, then reports how popular each act really is."

120b: k=1

See: A balcony overlooks a grand plaza. Three scouts sit stacked along the left railing, each wearing a distinctive color sash. Below, a bustling plaza paved in a single uniform color hosts a fair — small booths and troupes scattered about, each group wearing matching costumes of various colors. Some troupe colors match a scout's sash; others are uninvited performers who snuck in.

Do: Each scout leans over the railing and counts how many separate troupes in the plaza share their sash color — not the number of performers, but the number of distinct groups. The scout then unfurls a banner along the balcony railing, stretching it one tile for each troupe counted. A scout who finds no matching troupes pulls their banner down entirely and disappears. Meanwhile, down in the plaza, security sweeps through and escorts out every uninvited troupe — any group whose costume color doesn't match any scout's sash is absorbed back into the plaza floor, leaving only the scouted acts standing.

Grid: The balcony fills the upper floors (the top seven rows), a quiet black space where scouts sit only on the odd-numbered landings along the leftmost pillar. A solid colored cornice separates balcony from plaza. The plaza stretches below — thirteen rows paved in the cornice's color, with small clustered acts dotted throughout.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (12523 chars)

We need to infer rule from example. Input grid 20 rows? Let's count: example input has 20 rows? Actually given 20 rows (top 7 rows balcony, then a row of cornice? Let's examine.

In example 1 input: first 7 rows have single colored cells at left column: rows 2 R, 4 A, 6 Y. Others empty. Row 8 is G G G ... (full row of G). That's the cornice? Actually description: balcony top seven rows, quiet black space where scouts sit only on odd-numbered landings along leftmost pillar. So scouts are at rows 2,4,6 (odd-numbered landings? maybe rows 2,4,6). They have colors R, A, Y. Then row 8 is cornice of uniform color (G). Then rows 9-20 are plaza with groups of colors.

In output: the scouts rows have been extended horizontally: row2 now has R repeated 4 times then rest empty. Row4 A repeated 2 then rest empty. Row6 Y unchanged? Actually row6 Y . . . . . . . . . (no extension). Wait row6 Y unchanged because maybe no matching groups? Let's see.

In plaza, groups: we need to count distinct groups of each scout color. Then extend banner length equal to number of groups. Also remove groups whose color not matching any scout.

Let's analyze example.

Scouts colors: R, A, Y.

Plaza rows (starting row9?). Let's list groups in input.

Rows 9-20 (12 rows) after cornice row8.

I'll write grid of those rows (10 columns). Row9: G B B G G G R G A G
Row10: G B B G G G R G A G
Row11: G G G G G G R G A G
Row12: G G R R G G G G A G
Row13: G G G G G G G G G G
Row14: G G A A G G G G G G
Row15: G G A A G Y Y Y G G
Row16: G G G G G G G G G G
Row17: G R R G G G G G G G
Row18: G R R G G G G R R G
Row19: G G G G G G G G G G
Row20: G G G G G G G G G G

Now groups: groups are contiguous clusters of same color (connected orthogonally?). Likely groups of same color that are adjacent (4-direction). Need to count distinct groups of each scout color.

Scout R: find groups of red (R) in plaza. Let's locate R cells.

Rows9-10: column7 is R (positions (9,7) and (10,7)). They are vertically adjacent, forming one group? Also column? Row11 column7 R. Row12 column3-4 are R R (two adjacent). Row17 column2-3 are R R. Row18 column2-3 are R R, also column8-9 are R R. So groups:

Group1: the vertical column at col7 rows9-11 (three cells). That's one group.

Group2: row12 col3-4 (two cells) adjacent horizontally, separate from above (since row11 col3 is G). So second group.

Group3: rows17-18 col2-3 (a 2x2 block? rows17 col2-3 R R, row18 col2-3 R R). That's one group.

Group4: row18 col8-9 R R (two cells) separate.

Thus total 4 groups of R.

Scout A (azure) color A: locate A cells.

Rows9-11 column9? Actually column9 is A in rows9-11. That's vertical group of three cells (col9 rows9-11). That's one group.

Row12 column9 A (single) adjacent vertically to above? row11 col9 A, so actually continues: row9-12 col9 A forms a vertical group of 4 cells. So group1.

Row14 columns3-4 are A A (row14 col3-4). That's separate group2.

Row15 columns3-4 A A (row15 col3-4) adjacent vertically to row14 col3-4, so th

... [9544 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.01)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 965: "Colored stars cast pillars of light, and their constellation ... [38 chars truncated]" k=1 alts: 8/15 f5b8619d

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input is a rectangular grid with a black (.) background containing a small number of colored cells (all the same non-black color) scattered at various positions. Some columns contain one or more colored cells, while other columns contain no colored cells at all.

Do: 1. Identify which columns contain at least one colored cell. These are 'active' columns. 2. For each active column, fill every cell in that column with azure (A), then restore the original colored cells at their original row positions. 3. Leave all inactive columns (those with no colored cells) entirely black (.). This produces a transformed tile the same size as the input. 4. Tile this transformed grid in a 2×2 arrangement: place the tile in the top-left, top-right, bottom-left, and bottom-right, producing an output whose height is twice the input height and whose width is twice the input width.

Grid: The input grid can be any rectangular dimensions. The output grid is exactly twice the width and twice the height of the input, formed by repeating the transformed tile in a 2×2 pattern. The background color is black (.). Azure (A) is used exclusively as the fill color for activated columns and does not appear in the input.

Figurative Descriptions (16 variants)

"Colored stars cast pillars of light, and their constellation is printed on four panes of a window."

120b: k=1

See: A dark night sky dotted with a few colored stars. Some vertical lanes of the sky hold stars; others are completely empty and dark.

Do: Each star casts a pillar of azure light that floods its entire vertical lane from horizon to horizon, like a searchlight pointed straight up and down. The stars themselves burn through the glow, keeping their original color. Lanes with no stars remain pitch dark — the light has no source there. Once the illuminated sky is complete, it is photographed and printed identically onto all four panes of a divided window, two across and two down.

Grid: The night sky can be any size. The window is always twice as tall and twice as wide as the original sky, perfectly tiling four identical prints. Black is the darkness of empty space, and azure is the glow that only stars can summon — it never appears on its own.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (4140 chars)

We need to infer rule. Input grid: black '.' and maybe other colors (stars). In example, input has '.' and X (grey). Output is 2x size? Input size 6x6? Actually input lines: ". X . . . ." (6 columns) and 6 rows. Output appears 12x12 (twice dimensions). Indeed output is 12 rows each 12 columns. The rule: For each column (vertical lane), if any star (non-black?) present, then entire column in output becomes azure A, except original star cells retain their original color. Also the output is tiled 2x2 copies of this transformed grid? Let's examine: Input column 2 (index1) has X at row0 and row4? Actually input rows: row0: . X . . . . ; row4: X . . . . X. So column1 has X at row0, column0 has X at row4, column5 has X at row4. So stars are X (grey). In output, we see many A's and X's. Let's see output first row: "A X . . . A A X . . . A". That's 12 columns. It seems they tiled 2x2: top-left 6x6 transformed, top-right same, bottom-left same, bottom-right same. Let's verify: The transformed 6x6 would have azure A in columns that have any X in original, but keep X where originally. Original columns with X: col0 (row4), col1 (row0), col5 (row4). So transformed grid should have A in those columns for all rows, except at original X positions keep X. Other columns remain black '.'? Let's construct transformed 6x6: rows 0-5, cols0-5.

Initialize all '.'.

For each column with any X, set entire column to A. Then set original X cells back to X.

Columns with X: 0,1,5.

Thus column0 all A, but row4 col0 should be X. Column1 all A, but row0 col1 X. Column5 all A, but row4 col5 X.

Result grid:

Row0: col0 A, col1 X, col2 ., col3 ., col4 ., col5 A => "A X . . . A"
Row1: A A . . . A
Row2: A A . . . A
Row3: A A . . . A
Row4: X A . . . X
Row5: A A . . . A

That matches the top-left 6x6 of output? Let's check output rows 0-5 columns 0-5: Row0 "A X . . . A" yes. Row1 "A A . . . A" yes. Row2 same, Row3 same, Row4 "X A . . . X" yes, Row5 "A A . . . A". Good.

Then output repeats this block horizontally and vertically to make 12x12. Indeed rows 0-5 columns 6-11 repeat same pattern. Rows 6-11 repeat same as rows0-5.

Thus rule: Identify columns that contain any non-black (colored) cell (any color other than .). Fill entire column with azure A, but keep original colored cells unchanged. Then tile the resulting transformed grid 2x2 (duplicate both horizontally and vertically) to produce output of double dimensions.

Now test input: 4x4 grid.

Input:
Row0: . . G .
Row1: . . . .
Row2: . . . G
Row3: G . . .

Colors: G (green) at (0,2), (2,3), (3,0). So columns with any colored cell: column0 (row3), column2 (row0), column3 (row2). Column1 has none.

Thus transformed 4x4: set columns 0,2,3 to A, except keep G at original positions.

Construct:

Initialize all '.'.

Column0 all A, but row3 col0 = G (original). Column2 all A, but row0 col2 = G. Column3 all A, but row2 col3 = G.

Result rows:

Row0: col0 A, col1 ., col2 G, col3 A => "A . G A"
Row1: col0 A, col1 ., col2 A, col3 

... [1167 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.03)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 966: "A rubber stamp carved around a missing center is pressed whe ... [28 chars truncated]" k=2 alts: 4/15 f5c89df1

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains a pattern made of three colors: azure (A) cells forming a shape, a single green (G) cell at the center of that shape, and one or more red (R) cells positioned among or adjacent to the azure cells. The azure cells surrounding the green cell define a template shape. The red cells mark target locations where copies of that template should be placed.

Do: 1. Locate the single green (G) cell — this is the origin of the template. 2. Identify all azure (A) cells and compute each one's offset relative to the green cell (row_offset, col_offset). This set of offsets defines the template shape. 3. Locate all red (R) cells — these are the stamp targets. 4. Start with a blank grid (all black/background). 5. For each red cell, stamp a copy of the template: for every offset in the template, place an azure cell at (red_row + offset_row, red_col + offset_col), provided it falls within the grid bounds. 6. The original pattern (all G, A, and R cells) is removed — only the stamped copies remain.

Grid: The grid dimensions match the input (typically 13×13). The background is black (.). The output contains only azure (A) cells on a black background — no green or red cells remain. The coordinate system is row, column with (0,0) at top-left.

Figurative Descriptions (16 variants)

"A rubber stamp carved around a missing center is pressed wherever ink drops have fallen."

120b: k=2

See: On a dark desk sits a rubber stamp carved from azure material. At its heart, where the handle meets the block, is a single green pivot point — the grip you hold it by. Scattered nearby are red ink drops splattered on the desk, each marking a spot where someone wants an impression.

Do: Lift the stamp off the desk by its green handle, removing it entirely from view. Now press the stamp down once onto each red ink drop — the red vanishes beneath the impression, and the azure carved pattern appears around where the drop was. The original stamp and all the ink drops disappear, leaving only the fresh impressions behind on the dark surface.

Grid: The desk is a dark field. Each stamped impression is rendered in azure. Nothing remains of the original stamp or the red ink drops — only the ghost-impressions left behind, perfectly reproducing the stamp's carved shape at every spot where ink once sat.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (10074 chars)

We need to understand the puzzle. The metaphor: a stamp (azure) with a green pivot point (handle). Red ink drops are positions where we need to stamp. The stamp shape is the azure pattern (A) with a green pivot point (G) at its handle. In input, we have a stamp shape (A's and a G) somewhere, plus red R's scattered. The rule: remove the original stamp and red drops, and place copies of the stamp (azure shape) centered at each red drop location, but with the green pivot point removed? Wait: In output examples, we see only azure A's, no G. Indeed output has only A's, no G. The original stamp (including G) disappears. The red R's disappear. The azure pattern appears at each location where a red R was, but the pattern is the same as the original stamp shape (the arrangement of A's relative to the G pivot). However, the output shows A's placed at positions that correspond to the original stamp shape but shifted so that the G (green) aligns with the original R location? Let's verify.

Example 1 Input: Let's map coordinates. We'll index rows 0-12, columns 0-12.

Input has A's forming a plus shape? Let's list coordinates of A's and G.

Rows:

Row3: col6 = A (since row index 3, column 6). Actually row3: ". . . . . . A . . . . . .". So A at (3,6).

Row4: ". . . . . A A A . . . . .": A at (4,5),(4,6),(4,7).

Row5: ". . . R A . G . A R . . .": A at (5,4) and (5,8). G at (5,6). R at (5,3) and (5,9).

Row6: ". . . . . A A A . . . . .": A at (6,5),(6,6),(6,7).

Row7: ". . . . . . A . . . . . .": A at (7,6).

So the stamp shape is a cross of A's with a G at center? Let's see shape: coordinates relative to G at (5,6). Offsets: A at (3,6) offset (-2,0). Row4 A's at (4,5) (-1,-1), (4,6) (-1,0), (4,7) (-1,1). Row5 A at (5,4) (0,-2) and (5,8) (0,2). Row6 A's at (6,5) (1,-1), (6,6) (1,0), (6,7) (1,1). Row7 A at (7,6) (2,0). So shape is a diamond? Actually it's like a plus with arms length 2 horizontally and vertically, plus the four diagonal positions at distance sqrt(2) (the ones at (-1,-1), (-1,1), (1,-1), (1,1)). So shape includes 13 cells: center G, plus 12 A's? Wait count: A's: row3 (1), row4 (3), row5 (2), row6 (3), row7 (1) =10 A's. plus G =1 => total 11 cells. Actually we have 10 A's.

Now output Example1: The output shows A's at positions where the stamp was placed at each red R location. There were two R's at (5,3) and (5,9). In output, we see A's forming the shape centered at those positions? Let's check output coordinates.

Output rows:

Row3: ". . . A . . . . . A . . .": A at (3,3) and (3,9). That's two A's.

Row4: ". . A A A . . . A A A . .": A at (4,2),(4,3),(4,4) and (4,8),(4,9),(4,10).

Row5: ". A . . . A . A . . . A .": A at (5,1),(5,5),(5,7),(5,11).

Row6: ". . A A A . . . A A A . .": A at (6,2),(6,3),(6,4) and (6,8),(6,9),(6,10).

Row7: ". . . A . . . . . A . . .": A at (7,3) and (7,9).

These correspond to two copies of the original A pattern, shifted so that the G (center) would be at the original R positions (5,3) and (5,9). Let's verify: For 

... [7107 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

No grid

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 971: "A popularity contest among rare wildflowers in a vast meadow" k=1 alts: 8/15 f8b3ba0a

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid is composed of 2×2 colored blocks arranged in a regular grid pattern, separated by single-cell-wide black (.) borders. One color serves as the dominant background, appearing far more frequently than the others. The remaining minority colors each appear in one or more 2×2 block positions scattered among the background blocks.

Do: 1. Identify all distinct 2×2 colored blocks by scanning the grid at every position where blocks occur (they are spaced 3 cells apart both horizontally and vertically, starting at row 1, column 1). 2. Count the occurrences of each color across all block positions. 3. The color with the highest count is the background color — set it aside. 4. For each remaining minority color, record its count. 5. Sort the minority colors in descending order by count (most frequent first). 6. Output each minority color as a single row, one per line, from most frequent to least frequent.

Grid: The output grid is a single column with one row per minority color (typically 3 rows, 1 column). Each cell contains the color code of a minority color, ordered from most to least frequent. The input grid dimensions vary but always follow the pattern (2k+1) rows by (2m+1) columns, where k and m reflect the number of block rows and columns, with single-cell black borders separating all 2×2 blocks.

Figurative Descriptions (16 variants)

"A popularity contest among rare wildflowers in a vast meadow"

120b: k=1

See: A meadow stretches before you, tiled with patches of flowers. One species dominates the landscape so thoroughly it has become the grass itself — the wallpaper of the world. But scattered among this sea of sameness, a few rare wildflowers have taken root, their colors defiant against the monotony. Each flower bed is a small square plot, and the black soil paths between them form a quiet grid.

Do: The meadow keeper conducts an annual census of the rare species. She walks every plot, ignoring the common ground-cover entirely — it is invisible to her practiced eye. For each rare bloom she encounters, she places a tally mark in her notebook. When the walk is done, she ranks the rare species from most abundant to least, crowning the one closest to overtaking the meadow as the first name on her list. The rarest flower, barely clinging to existence, goes last.

Grid: The keeper's final report is a single column in her ledger — one flower color per line, written top to bottom from the most commonly spotted rarity down to the loneliest bloom in the field.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6015 chars)

We need to infer rule from example. Input grid has many rows, with pattern: black '.' as background, and colored squares appear in 2x2 blocks? Let's examine Example 1 input (provided). Let's reconstruct mentally.

Rows (I'll index):
Row0: . . . . . . . . . .
Row1: . G G . G G . G G .
Row2: . . . . . . . . . .
Row3: . B B . G G . G G .
Row4: . . . . . . . . . .
Row5: . G G . A A . G G .
Row6: . . . . . . . . . .
Row7: . G G . A A . G G .
Row8: . . . . . . . . . .
Row9: . G G . R R . R R .
Row10: . . . . . . . . . .
Row11: . R R . G G . G G .
Row12: . . . . . . . . . .

Output:
R
A
B

So output is a column listing colors: R, A, B (top to bottom). According to description: count rare species (non-dominant). The dominant species is the "wallpaper" i.e., the most common color? In example, which color is most frequent? Let's count occurrences of each color in input.

We have colors: G, B, A, R. Let's count approximate: Each 2x2 block appears multiple times. Let's list blocks:

Row1 has G G pairs at columns 1-2 and 4-5 and 7-8? Actually pattern . G G . G G . G G . So three 2x2? Actually each block is 2 cells horizontally, but vertical? The pattern repeats every other row with blank rows between. So each block is a 2x2 square of same color? Let's see rows 1 and 3 maybe form vertical pairs? Row1 has G G at positions (1,2), (4,5), (7,8). Row3 has B B at (1,2), G G at (4,5), G G at (7,8). So at columns 1-2, rows1-3 we have G on row1, B on row3, not same. So maybe each "plot" is a 2x2 block formed by two rows: rows1-2? But row2 is all dots. So maybe each plot is a 2x2 block formed by rows1 and 3? Not contiguous.

Actually pattern: there are rows with colors, separated by empty rows. So each "plot" is a 1x2 horizontal pair? The description says "small square plot". Could be 2x2 but due to empty rows, they are separated.

Let's examine: Row1 has G G . G G . G G . (positions). Row3 has B B . G G . G G . So there are three columns of pairs: first pair (col1-2) has G on row1, B on row3. So maybe each plot is a 2x2 formed by rows1 and 3 (with a blank row between). That would be a square of size 2 rows (with a blank row considered part of grid but not part of plot). So each plot consists of two colored rows separated by a blank row. So each plot is a 2x2 block of same color? Not exactly.

But maybe the rule: Count each distinct color that appears in at least one 2x2 block (i.e., a "flower") and ignore the most common color (the "grass") which is considered background. Then output the remaining colors sorted by frequency descending.

In example, output includes R, A, B. So G is omitted (most common). Let's verify counts: G appears many times (lots of G blocks). R appears maybe 2 blocks (row9 and row11). A appears 2 blocks (rows5 and7). B appears 1 block (row3). So order: R (most among rare), A (next), B (least). That matches output.

Thus rule: Identify the most frequent color (excluding black '.'), treat it as background and ignore it. Then list other colors sorted

... [3036 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.20)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 977: "A tapestry weaver mends a moth-eaten hole by reading the rhy ... [29 chars truncated]" k=1 alts: 5/15 f9012d9b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains a pattern of colored cells that tiles periodically in both rows and columns, forming a repeating 2D mosaic. A rectangular region of the grid has been replaced with black (.) cells, creating a 'hole' in the otherwise regular tiling pattern. The non-black portion of the grid is fully consistent with a single repeating tile.

Do: 1. Identify the repeating tile by finding the row period and column period of the non-black cells. Compare rows and columns in the intact (non-black) region to determine the smallest periods P_r (row period) and P_c (column period) such that the pattern repeats every P_r rows and every P_c columns. 2. Extract one copy of the repeating tile: a P_r × P_c subgrid from any fully intact region of the grid. 3. Locate the rectangular region of black (.) cells. 4. For each black cell at position (r, c), compute the tile value as tile[r mod P_r][c mod P_c]. 5. Output only the rectangular block of restored values that replaces the black cells, matching the dimensions of the hole.

Grid: The grid dimensions vary across instances. The grid is composed entirely of colored cells arranged in a periodic tiling pattern, with one contiguous rectangular region of black (.) cells representing the missing portion. The output grid has the same dimensions as the missing rectangular region and contains the tile values that fill the hole.

Figurative Descriptions (16 variants)

"A tapestry weaver mends a moth-eaten hole by reading the rhythm of the surviving threads."

120b: k=1

See: A grand tapestry hangs on the wall, its colored threads woven in a endlessly repeating motif — a small melodic phrase played over and over across rows and columns. But moths have eaten through one corner, leaving a rectangle of bare black canvas where the pattern should continue. The surviving fabric hums with perfect regularity, every stitch loyal to the same small refrain.

Do: Listen to the tapestry's song — find the shortest melodic phrase that, when looped, reproduces every intact thread. Once you know the refrain by heart, hold it up to the moth-eaten hole and let the melody fill in what the moths stole. Sing only the missing notes — the patch that makes the tapestry whole again.

Grid: The tapestry varies in size but is always woven from a small repeating motif of colored threads tiling seamlessly in both directions. The moth-eaten hole is a contiguous rectangle of bare black canvas. Your repair patch matches the hole's dimensions exactly, carrying only the colors the refrain demands at each position.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (6305 chars)

We need to infer the puzzle: given a grid with a repeating pattern and a rectangular hole of '.' (black). Need to output the patch that fills the hole: the colors that should be there according to the repeating motif.

In example 1: Input grid:
A M . M
M A M A
A M A M
M A M A

Output: A

So the hole is a single '.' at position (row1,col3) (0-index?). The repeating motif appears to be a 2x2 pattern? Let's examine grid without hole.

Rows:
1: A M . M
2: M A M A
3: A M A M
4: M A M A

Looks like a checkerboard of A and M alternating both rows and columns. Pattern: A M
M A repeating. Indeed a 2x2 tile. So the missing cell should be M? Wait pattern: coordinates (r,c) parity: if (r+c) even => A? Let's test: (0,0) A (even) => A. (0,1) M (odd) => M. (0,2) missing should be A? Actually (0+2)=2 even => A. But output is A. Yes matches. So output is A. Good.

Thus rule: Determine smallest repeating tile (both dimensions) that tiles the grid (ignoring '.'), then fill the '.' rectangle with appropriate colors from that tile.

Now test input grid:

We have 7 columns? Let's write rows:

Row0: A B A A . . .
Row1: B A A B . . .
Row2: A A B A . . .
Row3: A B A A B A A
Row4: B A A B A A B
Row5: A A B A A B A
Row6: A B A A B A A

So grid 7x7? Actually rows 0-6 (7 rows) and columns 0-6 (7 columns). The hole is a rectangle of '.' in top right corner? Let's locate '.' positions: rows 0-2 have '.' at columns 4,5,6? Row0: columns 4,5,6 are '.' '.' '.' (three). Row1: columns 4,5,6 '.' '.' '.' . Row2: columns 4,5,6 '.' '.' '.' . So a 3x3 rectangle at rows 0-2, cols 4-6.

Goal: find repeating motif that tiles entire grid (including hole area). The rest of grid (non '.' cells) should be consistent with some small pattern.

We need to find smallest tile dimensions that repeat both horizontally and vertically.

Let's extract the known part: rows 3-6 are fully filled (no holes). Let's write them:

Row3 (index3): A B A A B A A
Row4: B A A B A A B
Row5: A A B A A B A
Row6: A B A A B A A

Observe pattern maybe period 4? Let's examine columns.

Write grid as matrix:

r\c:0 1 2 3 4 5 6
0: A B A A . . .
1: B A A B . . .
2: A A B A . . .
3: A B A A B A A
4: B A A B A A B
5: A A B A A B A
6: A B A A B A A

We can try to see if pattern is 4x4? Let's look at rows 3-6 maybe repeat rows 0-3? Row3 equals row0? Row0 is A B A A ... Row3 is A B A A B A A (extra). Not same because row0 has '.' after col3. But ignoring hole, first 4 columns of row3: A B A A matches row0 first 4 columns A B A A. Row4 first 4 columns: B A A B matches row1 first 4 columns B A A B. Row5 first 4 columns: A A B A matches row2 first 4 columns A A B A. Row6 first 4 columns: A B A A matches row0 first 4 columns again. So seems pattern repeats every 3 rows? Actually rows 0-2 pattern repeats in rows 3-5? Row3 matches row0, row4 matches row1, row5 matches row2, row6 matches row0 again. So vertical period 3? Let's check: rows 0,3,6 are same (first 4 columns). rows1,4 same. rows2,5 same. So period 3 rows.

Now hori

... [3383 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.11)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 982: "A mirror placed at the corner of a room doubles the world, a ... [36 chars truncated]" k=2 alts: 3/15 fb791726

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains a black background with one or more vertical pairs of same-colored non-black cells. Each pair consists of two cells of the same color in the same column, separated by exactly one row between them. The row between the two cells of each pair is entirely black. All pairs use the same color.

Do: 1. Create an output grid with double the width and double the height of the input, filled entirely with black cells. 2. Copy the entire input grid into the top-left quadrant of the output (rows 0 to H-1, columns 0 to W-1, where H and W are the input height and width). 3. For each vertical pair of colored cells, fill the entire row between them (across the full output width) with green (G). 4. Create a diagonal copy: copy the input grid again into the bottom-right quadrant (rows H to 2H-1, columns W to 2W-1). 5. For each vertical pair in this copied region, again fill the entire row between them (across the full output width) with green (G).

Grid: The output grid is exactly twice the height and twice the width of the input grid. The background is black. Green separator rows span the full width of the output grid. The input content appears in the top-left quadrant, and a shifted copy appears in the bottom-right quadrant, forming a diagonal tiling pattern.

Figurative Descriptions (16 variants)

"A mirror placed at the corner of a room doubles the world, and every doorway gets a welcome mat."

120b: k=2

See: On a dark stage, pairs of colored lanterns hang in vertical columns, each pair separated by a single dark gap—like eyes peering through a slot. They all glow the same hue, watching and waiting.

Do: The stage expands diagonally: the original scene sits in one corner, and its echo appears in the far opposite corner, as if reflected in a mirror placed at the room's center. Wherever two lanterns stare at each other across a dark gap, a green carpet unrolls between them—not just beneath their feet, but stretching wall to wall across the entire widened stage. Each pair of watchful eyes earns its own green bridge, in both the original scene and its echo.

Grid: The new stage is twice as tall and twice as wide as the original, mostly darkness. The original performance occupies one corner, its twin occupies the diagonal opposite corner, and vivid green runways cut horizontally across the full width wherever paired lanterns face each other.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (10139 chars)

We need to understand the transformation. Let's parse description: Input grid with colored lanterns (non-black?) placed in vertical columns, each pair separated by a single dark gap. They all glow same hue, watching and waiting. So input likely has pairs of same color in same column with a dot between them? Let's examine examples.

Example1 Input (3x3):
Row1: . A .
Row2: . . .
Row3: . A .
So there are two A's in column 2? Actually column indices: col1 ., col2 A, col3 .; row2 all ., row3 col2 A. So A's are vertically aligned with a gap of one row between them (row2). So they are a pair separated by a single dark gap. They are same hue (A). Output is 6x6 (twice dimensions). Let's view output.

Output rows (6):
1: . A . . . .
2: G G G G G G
3: . A . . . .
4: . . . . A .
5: G G G G G G
6: . . . . A .

Let's map coordinates. The original 3x3 placed in top-left corner? Let's see top-left 3x3 of output:
Rows1-3, cols1-3:
Row1: . A .
Row2: G G G
Row3: . A .
But output row2 col1-3 are G G G, not . . . . So they inserted a full row of G across entire width (6) at row2 (the row between the two A's). Indeed original had gap row2 (all .). In output, that row becomes all G across width. Also original top-left 3x3 appears unchanged except that row2 is now G row across width. Also there is a copy (echo) placed in opposite corner (bottom-right). Let's examine bottom-right 3x3 (rows4-6, cols4-6):
Rows4-6, cols4-6:
Row4: . A .
Row5: G G G
Row6: . A .
Matches same pattern. Indeed output rows4-6, cols4-6 replicate original pattern (with G row). Also there are other cells: rows4-6, cols1-3 and rows1-3, cols4-6 are mostly . except some A's? Let's see output row4: ". . . . A ." Actually row4 col5 is A (since original A at column2 row1? Wait). Row4 is ". . . . A .": So there is an A at column5 (col5). That's part of the echo? In echo, original A's are at column5? Let's compute: Echo placed at bottom-right corner offset by (original rows, cols). Original size 3x3, output size 6x6. So echo origin at row4, col4 (since 3+1). In echo, the A's should be at (row4+0, col4+1) = (4,5) and (row4+2, col4+1) = (6,5). Indeed row4 col5 A, row6 col5 A. Good. So echo works.

Now also there is a G row across full width at row5 (the row between the two A's in echo). Indeed row5 is all G.

Thus rule: For each column where there are two same-colored non-blank cells separated by exactly one blank cell (i.e., pattern color, blank, same color vertically), we create a full horizontal line of G across entire output grid at the row corresponding to the blank row (i.e., the middle row). Also we duplicate the original pattern into opposite corner (mirror across both axes). The output grid is double size (2x width and height). The original pattern placed top-left, the copy placed bottom-right. The rest is filled with . except the G rows (full width) at appropriate rows (both for original and copy). Also maybe there are multiple columns with such pairs; each pair yields a G row (maybe multiple rows). 

... [7142 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.26)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 983: "A pair of scissors snips along the dotted line, and the scra ... [19 chars truncated]" k=1 alts: 1/15 fbf15a0b

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Literal Description

See: The input grid has an azure (A) background containing one or more rectangular regions of colored pattern (repeating motifs of a non-azure color on every other cell). Exactly two grey (X) marker cells appear on one edge of the grid, positioned two cells apart from each other along that edge. The X markers always sit on the outermost row or column.

Do: 1. Locate the two X marker cells. They will be on the same edge (top row, bottom row, left column, or right column), spaced 2 cells apart. 2. Determine the axis the X markers define: if they are on the top or bottom row, the axis is vertical (columns); if on the left or right column, the axis is horizontal (rows). 3. Halve the grid along that axis, splitting it into two equal parts. If the axis is vertical, split into a left half and right half; if horizontal, split into a top half and bottom half. 4. Keep the half that contains the X markers; discard the other half. 5. In the kept half, replace every X cell with azure (A). The result is the output grid.

Grid: The input grid has even dimensions along the axis being halved (so it splits cleanly in two). The output grid is exactly half the size of the input along the halved axis and the same size along the other axis. The background color is azure (A). The colored pattern regions and their internal spacing are preserved unchanged in the output, except that X markers become A.

Figurative Descriptions (16 variants)

"A pair of scissors snips along the dotted line, and the scrap paper falls away."

120b: k=1

See: An azure canvas stretches out, decorated with repeating tapestry patterns woven in various colors. Somewhere along one edge, two grey pins have been pressed into the frame — a perforation mark, like the tear-here dots on a coupon.

Do: The two pins mark where the scissors should cut. They whisper which direction to slice — across or down — like a dotted line on wrapping paper. Cut the canvas in half along that line. The half with the pins is the keeper; the other half flutters to the floor like a trimmed margin. Once you've kept your piece, pull out the pins and patch the tiny holes with azure, leaving no trace they were ever there.

Grid: The canvas is always sized so the scissors cut cleanly in two — no ragged edges. The kept piece is exactly half the original in one direction, unchanged in the other. The woven tapestry patterns survive the cut perfectly intact; only the pin-holes vanish.

gpt-oss-120b — MARC valid, solved with 1 example

Input

→

Prediction

vs

Expected

Reasoning (10813 chars)

We need to infer rule. From description: canvas azure (A) background, repeating tapestry patterns in various colors. Two grey pins (X) mark where to cut. They indicate direction across or down. Cut canvas in half along that line. The half with the pins is kept; other half discarded. Then remove pins (replace with azure A). So output is half of input (the half containing X's) with X replaced by A.

We need to determine orientation: cut line is across (horizontal) or down (vertical) depending on arrangement of X's? In example, input grid size? Let's parse example 1.

Input appears 20 rows? Let's count rows: Provided lines: 
1: X A A A A A A A A A A A A A A A A A A A
2: A A A A A A A A A A A A A A A A A A A A
3: X B A B A B A B A B A B A B A B A B A B
4: A A A A A A A A A A A A A A A A A A A A
5: A B A B A B A B A B A B A B A B A B A B
6: A A A A A A A A A A A A A A A A A A A A
7: A B A B A B A B A B A B A B A B A B A B
8: A B A B A B A B A B A B A B A B A B A B
9: A A A A A A A A A A A A A A A A A A A A
10: A B A B A B A B A B A B A B A B A B A B
11: A B A B A B A B A B A B A B A B A B A B
12: A A A A A A A A A A A A A A A A A A A A
13: A B A B A B A B A B A B A B A B A B A B
14: A A A A A A A A A A A A A A A A A A A A
15: A B A B A B A B A B A B A B A B A B A B
16: A A A A A A A A A A A A A A A A A A A A
17: A A A A A A A A A A A A A A A A A A A A
18: A A A A A A A A A A A A A A A A A A A A
19: A A A A A A A A A A A A A A A A A A A A
20: A A A A A A A A A A A A A A A A A A A A
21: A A A A A A A A A A A A A A A A A A A A

Actually there are 21 rows? Let's count lines: after row 20 maybe 21. Output has 10 rows? Output lines: 
1: A A ... (20 columns)
2: A A ...
3: A B A B...
4: A A...
5: A B...
6: A A...
7: A B...
8: A B...
9: A A...
10: A B...

So output is 10 rows, half of 20? Input maybe 20 rows, but we counted 21. Let's recount: maybe first line is row0? Let's count again with numbers:

1 X A A A A A A A A A A A A A A A A A A A
2 A A A A A A A A A A A A A A A A A A A A
3 X B A B A B A B A B A B A B A B A B A B
4 A A A A A A A A A A A A A A A A A A A A
5 A B A B A B A B A B A B A B A B A B A B
6 A A A A A A A A A A A A A A A A A A A A
7 A B A B A B A B A B A B A B A B A B A B
8 A B A B A B A B A B A B A B A B A B A B
9 A A A A A A A A A A A A A A A A A A A A
10 A B A B A B A B A B A B A B A B A B A B
11 A B A B A B A B A B A B A B A B A B A B
12 A A A A A A A A A A A A A A A A A A A A
13 A B A B A B A B A B A B A B A B A B A B
14 A A A A A A A A A A A A A A A A A A A A
15 A B A B A B A B A B A B A B A B A B A B
16 A A A A A A A A A A A A A A A A A A A A
17 A A A A A A A A A A A A A A A A A A A A
18 A A A A A A A A A A A A A A A A A A A A
19 A A A A A A A A A A A A A A A A A A A A
20 A A A A A A A A A A A A A A A A A A A A
21 A A A A A A A A A A A A A A A A A A A A

Actually there are 21 rows, but output has 10 rows. Maybe cut is vertical? Let's examine columns: Input width 20 columns. Output width also 20 columns. So cut is horizontal, halving row

... [7846 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.07)

Prediction

Expected

Figurative + 1 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 987: "A castle's walls absorb the color of the invading army's ban ... [5 chars truncated]" k=2 alts: 7/15 fcb5c309

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Literal Description

See: The input grid contains two hollow rectangles outlined in the same color (e.g., blue, red, or green), one larger and one smaller. Their borders are drawn as single-cell-thick lines of that color. Scattered across the grid, both outside and inside the rectangles, are dots of a second, different color. The rest of the grid is black (background). The interior dots inside the larger rectangle occupy specific positions within its interior.

Do: 1. Identify the two hollow rectangles in the grid. They share the same border color. One is larger than the other. 2. Identify the second color — the color of the scattered dots found outside the rectangles. 3. Extract the larger rectangle (including its border and interior). 4. Replace every cell that uses the rectangle's border color with the scattered-dot color. Interior cells that are black remain black; interior cells that already contain the scattered-dot color remain that color. 5. Output the resulting rectangle as the answer grid.

Grid: The output grid has the same dimensions as the larger rectangle found in the input (height and width matching its bounding box, inclusive of borders). The border of the output is entirely the scattered-dot color. Interior cells are either black (.) or the scattered-dot color, preserving the positions of any dots that appeared inside the larger rectangle in the input.

Figurative Descriptions (16 variants)

"A castle's walls absorb the color of the invading army's banners."

120b: k=2

See: Two fortresses stand on a dark battlefield, their walls built from the same stone color. One fortress is grand, the other modest. Scattered across the battlefield like campfires are sparks of a rival color — an invading force that has spread its banners everywhere, even slipping agents inside the fortress walls.

Do: The invaders conquer the larger fortress. Their banners replace every stone in its walls, claiming it entirely. The spies already inside hold their positions. Empty rooms remain empty — the conquest changes the flag, not the furniture. The smaller fortress is forgotten, lost to history.

Grid: The conquered fortress stands alone as the output — same shape, same size, but now every wall stone flies the invader's color. Inside, rooms that were empty stay dark, and any invader agents who had already infiltrated remain where they stood. The fortress's silhouette is unchanged, but its allegiance has shifted completely.

gpt-oss-120b — MARC valid, solved with 2 examples

Input

→

Prediction

vs

Expected

Reasoning (12404 chars)

We need to infer rule from examples.

We have metaphor: two fortresses built from same stone color. One larger, one smaller. Invaders of rival color spread everywhere, even inside fortress walls. Conquer larger fortress: replace every stone in its walls with invader's color, but keep interior empty cells unchanged, and any invader agents already inside remain. Smaller fortress is ignored.

Thus we need to identify two shapes (fortresses) made of same color (stone). Larger shape will be transformed: its border cells become invader color, interior stays same (empty or invader). The invader color is the other color that appears scattered across grid (like "sparks").

In examples, we need to see.

Example1 Input: many colors: G (green) appears as border? B (blue) appears as interior? Let's examine.

Input grid size? Looks like 13 rows? Actually list: lines show 13 rows? Let's count: lines from Input example1: first line ". . . . . . . . . . . . . G . ." that's 16 columns. There are 13 rows (since output is 6x7). The output is a shape of size 6x7 with G border and interior '.' and some G inside? Let's view output:

Output:
G G G G G G G
G . . G . . G
G . . . . . G
G G . G . . G
G . . . . . G
G G G G G G G

That's a shape of a fortress: border G, interior '.' except some internal G at positions (2,4) and (4,2?) Actually row2 col4 is G, row4 col2 is G? Let's map coordinates (1-index). Row1 all G. Row2: G . . G . . G (so G at col1,4,7). Row3: G . . . . . G (col1,7). Row4: G G . G . . G (col1,2,4,7). Row5: G . . . . . G (col1,7). Row6: all G.

So shape is like a rectangle with some internal G "walls"? Might correspond to original larger fortress border (stone color) being G? In input, G appears as scattered? Let's examine input: there are many B's forming a big block maybe representing fortress walls? Actually B appears as large rectangle of B's maybe the larger fortress. G appears as scattered points inside and border? Let's look: Input rows (I'll rewrite with coordinates). Row1: ... G at col14? Row2: G at col5. Row3: G at col4, then B's from col9-15? Row4: B at col9, G at col13? Row5: G at col2, G at col6, B at col9, B at col15? Row6: G at col7, B at col9, G at col11, G at col13, B at col15. Row7: B's at col2-8? Actually row7: ". B B B B B B . B B B B B B B ." So B's from col2-8, then dot, then B's col10-15? Row8: ". B . . . . B . . . . . . . . ." B at col2 and col8. Row9: ". B . . . G B . . . . . . . . ." B at col2, G at col6, B at col7. Row10: ". B . . . . B . . . . . . . . ." B at col2 and col8. Row11: "G B B B B B B . . . . G . . . ." B from col2-7, G at col1 and col12. Row12: "G . . . . . . G . . G G . . . ." G at col1,8,11,12.

So there is a large rectangle of B's roughly rows3-11 columns2-15 maybe? Let's see: B's form a shape maybe like a fortress walls of B (stone). G are scattered inside and also on border? The output shape is G border, interior '.' except some G internal (maybe original G that were inside). Indeed output shows G border (the st

... [9440 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.00)

Prediction

Expected

Figurative + 2 ex PASS (acc: 1.00)

Prediction

Expected

▶ Task 995: "A shooting star streaks across the night sky, its colored fr ... [27 chars truncated]" k=4 alts: 14/15 feca6190

Training Examples

Example 1

Input

→

Output

Example 2

Input

→

Output

Example 3

Input

→

Output

Example 4

Input

→

Output

Example 5

Input

→

Output

Literal Description

See: The input is a single row of 5 cells. Some cells are black (background) and some are non-black (colored). The non-black cells form a pattern embedded within the row, with black cells acting as spacers between or around them.

Do: 1. Count the number of non-black cells in the input row. Call this number N. 2. Create a square output grid of size (5×N) by (5×N), filled entirely with black cells. 3. Place the input row as the bottom row of the output grid, left-aligned (occupying the first 5 columns, with the remaining columns filled with black). 4. For each row above the bottom, shift the entire 5-cell input pattern one column to the right compared to the row below it. Cells that shift beyond the right edge of the output grid are discarded. Cells vacated on the left become black. 5. Continue this diagonal shifting upward until the top row, where the pattern has been shifted right by (5×N − 1) positions. Only the non-black cells that still fall within the grid bounds are drawn; all other cells remain black.

Grid: The output grid is always square. Its side length equals 5 times the number of non-black cells in the input row (e.g., 2 non-black cells → 10×10, 4 non-black cells → 20×20). The background color is black. The input's 5-cell pattern traces a diagonal stripe from the bottom-left area to the top-right corner, with each row offset one column to the right from the row below.

Figurative Descriptions (16 variants)

"A shooting star streaks across the night sky, its colored fragments trailing behind it."

120b: k=4

See: A handful of colored sparks sit loaded in a launch rail at the bottom of a dark night sky. The black cells are empty sky; the colored cells are the glowing fragments of a firework about to be launched diagonally upward.

Do: Light the fuse. The colored sparks launch from the bottom-left and streak toward the upper-right corner, leaving a diagonal trail across the darkness. Each moment the cluster of sparks drifts one step higher and one step to the right, painting a luminous streak across the sky. The brighter the firework — the more colored sparks it carries — the bigger the sky it needs to cross.

Grid: The night sky is a square canvas whose size grows with the brilliance of the firework — more sparks demand a grander stage. The streak runs from the bottom-left corner to the upper-right, a single glowing diagonal ribbon of color against a field of pure darkness.

gpt-oss-120b — MARC valid, solved with 4 examples

Input

→

Prediction

vs

Expected

Reasoning (9829 chars)

We need to infer rule. From examples, input is a single row? In examples, input appears as a row of 5 cells. Output is a square grid of size (input length *? maybe 5x5? Actually output size seems larger: Example1 input length 5, output is 15x15? Let's count rows: they show 15 rows, each row 15 columns. So output size = (input length *3)? 5*3=15. Example2 input length 5, output 5x5? Wait example2 output shows 5 rows, each 5 columns. Actually they show 5 rows. So size = input length? Example2 output is 5x5. Example3 input length 5, output 15x15 again. Example4 input length5, output 5x5. So pattern: sometimes output size 5, sometimes 15. What determines? Possibly number of non-black (colored) cells in input. Example1 input: ". W . A Y" has three colored cells (W, A, Y). Output size 15. Example2 input: ". . R . ." has one colored cell (R). Output size 5. Example3 input: "Y . M . A" has three colored cells (Y, M, A). Output size 15. Example4 input: ". Y . . ." has one colored cell (Y). Output size 5.

Thus size = 5 * number of colored cells? For 3 colored -> 15. For 1 colored ->5. So output dimension = 5 * count(colored). Good.

Now transformation: It seems to create diagonal from bottom-left to top-right, placing the colored sequence along that diagonal, repeated? Let's examine example2: Input sequence left to right: . . R . . . Only R at position 3 (0-index 2). Output 5x5 grid:

Row0: . . . . .
Row1: . . . . .
Row2: . . . . R
Row3: . . . R .
Row4: . . R . .

So diagonal from bottom-left (row4 col0) upward right? Actually bottom-left is row4 col0 (since rows top to bottom). The diagonal of R's appears as a line of length 3? Wait there are three R's placed? Let's list coordinates with R: (row2,col4), (row3,col3), (row4,col2). That's a diagonal of length 3, sloping down-left to up-right? Actually as row increases (downwards), column decreases. That's a diagonal from top-right to bottom-left. But description says from bottom-left to upper-right. However orientation may be reversed due to coordinate system. Let's see example1: Input colored sequence: W A Y (positions 2,4,5?). Actually input row: ". W . A Y" positions: col0 ., col1 W, col2 ., col3 A, col4 Y. Output 15x15 shows many rows. Let's examine first few rows: Row0: all . (15). Row1: ... W at column? Row1: ". . . . . . . . . . . . . . . W" So W at far right column 14. Row2: W at column13. Row3: W at col12, A at col13? Actually row3: ". . . . . . . . . . . . W . A" So W at col11? Wait need to count. Let's think pattern: It seems they are drawing the colored sequence along a diagonal from top-right to bottom-left? Let's examine bottom rows: last row (row14): ". W . A Y . . . . . . . . . ." So bottom row has W at col1, A at col3, Y at col4. That's same as input positions but shifted left? Actually bottom row corresponds to original input row placed at bottom-left? Let's see bottom row: columns: 0 .,1 W,2 .,3 A,4 Y, rest . . So bottom row replicates input row at left side. Then each row above shi

... [6868 chars truncated]

Test

Test Input

→

Expected Output

MARC Property (original clue)

Examples Only FAIL (acc: 0.82)

Prediction

Expected

Figurative + 4 ex PASS (acc: 1.00)

Prediction

Expected

MARC2 Puzzle Inspector March 26, 2026

▶ Task 6: "A sail catching the wind snaps taut against its mast" k=1 alts: 13/15 025d127b

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 35: "Seeds dropped in soil grow into sedimentary rock layers, eac ... [44 chars truncated]" k=1 alts: 14/15 0f63c0b9

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 45: "Small tribes without enough members to survive are absorbed ... [20 chars truncated]" k=1 alts: 2/15 12eac192

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 47: "Archaeologists piece together a shattered mosaic from fragme ... [36 chars truncated]" k=2 alts: 8/15 137eaa0f

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 58: "A banner unfurls along the flagpole that holds it" k=3 alts: 7/15 15696249

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 85: "A snake slithers through a window frame, its body swaying si ... [11 chars truncated]" k=4 alts: 7/15 1c56ad9f

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 100: "Finding the one bird that stayed with its flock while all ot ... [26 chars truncated]" k=1 alts: 0/15 1f85a75f

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 136: "Musical chairs at a costume party — everyone moves one ... [43 chars truncated]" k=1 alts: 8/15 2601afb7

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 141: "A popularity contest where colors line up for a podium photo ... [50 chars truncated]" k=1 alts: 2/15 2753e76c

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 151: "Lightning rods on a rooftop attract bolts from storms on the ... [9 chars truncated]" k=1 alts: 9/15 29700607

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 194: "A curtain divides two dancers, and the spotlight finds whoev ... [16 chars truncated]" k=3 alts: 14/15 34b99a2b

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 207: "A talent show where the most popular act wins the encore." k=1 alts: 9/15 39a8645d

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 218: "A living cell divides, turning itself inside-out and stretch ... [59 chars truncated]" k=2 alts: 6/15 3befdf3e

Training Examples

Literal Description

Figurative Descriptions (16 variants)

Test

MARC Property (original clue)

▶ Task 223: "Scattered seeds fall and pile into pyramids on the ground." k=1 alts: 5/15 3d6c6e23