NARC Inspector

NARC puzzles are designed so that neither the grid sequence nor the narrative clue is sufficient on its own — but together they uniquely determine the answer. The experiments below probe different dimensions of this narrative–visual interaction. Masking tests whether models can reconstruct hidden grids. Ordering tests whether narratives help recover temporal structure. Odd-one-out tests whether narratives help models recognize which grids belong together. Stances tests how the same visual pattern responds to different narrative framings.

How to read this page. Masking is the documented battery: every active puzzle, all current models, order and keyword ablations; a paper on it is under review (September 2026). Ordering, Odd-one-out and Stances are pilot designs from spring 2026; each states its own scope and dates below. Counts are per model. The NARC property is a relation between a puzzle and a model, so a model with more NARC puzzles is not a better or worse model, and this page is not a leaderboard.

Masking Ordering Odd-One-Out Stances

Stance experiment: Inspired by Daniel Dennett's three interpretive stances. The same grid sequence is paired with three different narratives: intentional (beliefs, desires, goals — e.g., "Red betrayed the truce"), design (functional rules — e.g., "When blue reaches column 1, red overrides"), and physical (step-by-step cell changes — e.g., "Cell (1,1) changes from 1 to 2"), and moral (evaluative reasons — an agent overrides an expected trajectory for ethical reasons, e.g., a soldier refusing an unlawful order). Each is tested via the standard 3-condition masking protocol.

I = Intentional    D = Design    P = Physical    M = Moral    N/8 = NARC count out of 8 models    G = grids only    N = narrative only    B = both

Pilot design. 57 stance groups, 8 models. Stance narratives are stored as narrative variants, so they run through the same battery as Masking (model results April 9, 2026 to September 17, 2026); the stance framing itself is a spring 2026 pilot and not part of the paper's claims.

The Coup d'Etat
I:5/7  D:7/8  P:8/8 
Same grids, different narratives, different results
The Ambush
I:1/7  D:7/8  P:8/8 
Same grids, different narratives, different results
The Crossing
M:3/8 
Same grids, different narratives, different results

57 stance groups · Same grids described three ways: intentional, design, physical

▶ Hunt 036
I: 5/7
Intentional (5/7 NARC)
Blue ventured into the gap between them. Red betrayed the truce and took what blue had just claimed for itself.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
▶ Hunt 037
I: 7/7
Intentional (7/7 NARC)
Blue made the first move into the neutral zone. Red retaliated by seizing the very ground blue had just won.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
▶ Hunt 038
I: 3/7
Intentional (3/7 NARC)
Blue claimed the center, but red wanted it more. Red took it by force, turning blue's gain into red's.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b × × ×
▶ Hunt 039
I: 5/7
Intentional (5/7 NARC)
They shared a border. Blue pushed into the middle first. Red answered by stealing the position blue had just planted.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × ✓ ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
▶ Hunt 040
I: 7/7
Intentional (7/7 NARC)
Blue dared to break the stalemate and grabbed the center. Red punished the move, overrunning the position before blue could hold it.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
▶ The Alliance
I: 3/7 D: 5/8 P: 6/8
Intentional (3/7 NARC)
Blue and green formed a pact against red. They squeezed red from both sides. Green finished the job, taking the last of red's ground while blue stepped back.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
Design (5/8 NARC)
Blue from row 0 takes (1,0). Green from row 2 takes (1,2). Then green overrides red at (1,1), changing it to green. Blue's cell at (1,0) then changes back: in the next step blue withdraws from (1,0) and it becomes green.
ModelGNB
gemma-4-26b ✓ × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ×
Physical (6/8 NARC)
Step 0: row 0 all 1, row 1 all 2, row 2 all 3. Step 1: (1,0) changes from 2 to 1, (1,2) changes from 2 to 3. Step 2: (1,1) changes from 2 to 3. Step 3: (1,0) changes from 1 to 3.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Ambush
I: 1/7 D: 7/8 P: 8/8
Intentional (1/7 NARC)
The advance party engaged the target head-on, but a hidden flanking force appeared behind the enemy to cut off retreat.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (7/8 NARC)
Blue expands rightward along row 1. Independently, new blue cells appear in column 2 at rows 0 and 2 when the row-1 expansion reaches column 1. Then blue fills column 1 from rows 0 and 2. Finally blue fills column 0 rows 0 and 2. Red at (1,2) never changes.
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: cell (1,0) is 1, cell (1,2) is 2, rest 0. Step 1: cell (1,1) changes to 1. Step 2: cells (0,2) and (2,2) change to 1. Step 3: cells (0,1) and (2,1) change to 1. Step 4: cells (0,0) and (2,0) change to 1.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Armistice
I: 2/7
Intentional (2/7 NARC)
The ceasefire held until blue crossed the line. Red retaliated by capturing blue's forward position.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super ✓ × ✓
qwen3.5-122b × × ×
qwen3.6-27b × × ×
▶ The Backstab
I: 2/7 D: 6/8 P: 6/8
Intentional (2/7 NARC)
Blue reached across the divide, but red turned on its neighbor and snatched the foothold blue had just established.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (6/8 NARC)
Blue extends to (0,1). Red overrides (0,1) from blue to red. Red then takes (1,1).
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ×
Physical (6/8 NARC)
Step 0: col 0 all 1, col 2 all 2, col 1 all 0. Step 1: (0,1) becomes 1. Step 2: (0,1) changes from 1 to 2. Step 3: (1,1) becomes 2.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Betrayal
I: 5/7 D: 7/8 P: 7/8
Intentional (5/7 NARC)
Blue invaded green's corner. Red, pretending to be an ally, swooped in and stole what blue had just conquered.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (7/8 NARC)
Blue at row 0 places a cell at (2,0) in green's row. Red then overrides (2,0), changing it from blue to red. Red continues taking (2,1) from green.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
Physical (7/8 NARC)
Step 0: (0,0)=1,(0,1)=1, row 2 all 3. Step 1: (2,0) becomes 1. Step 2: (2,0) changes from 1 to 2. Step 3: (2,1) changes from 3 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Bluff
I: 0/7 D: 2/8 P: 8/8
Intentional (0/7 NARC)
The spy disguised itself as blue to blend in. Once surrounded, it revealed its true color in the center of the middle row.
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ×
gpt-oss-120b × ✓ ✓
gpt-oss-20b × ✓ ×
nemotron-3-super × × ×
qwen3.5-122b × ✓ ✓
qwen3.6-27b × ✓ ×
Design (2/8 NARC)
Rows fill with blue from top to bottom. After all cells become blue, the center cell of the grid changes to red. In subsequent steps, all cells in the center row convert to red.
ModelGNB
gemma-4-26b × ✓ ×
gemma-4-31b × ✓ ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: cells (0,0) through (0,2) set to value 1. Step 1: cells (1,0) through (1,2) set to value 1. Step 2: cells (2,0) through (2,2) set to value 1, then cell (1,1) changes from value 1 to value 2. Step 3: cells (1,0) and (1,2) change from value 1 to value 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Border War
I: 5/7 D: 0/8 P: 6/8
Intentional (5/7 NARC)
Green and red split the border. Green tested the gap first, but red betrayed the agreement and seized what green had just claimed.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (0/8 NARC)
Green fills column 0, red fills column 2. Column 1 is contested. Green claims the center cell (1,1) first. Then red overrides green's cell at (1,1), changing it from green (value 3) to red (value 2). So the masked grid has column 0 all green, column 2 all red, and (1,1) is now red. Specifically: [[3,0,2],[3,2,2],[3,0,2]].
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b × ✓ ✓
gpt-oss-20b × ✓ ✓
nemotron-3-super × ✓ ✓
qwen3.5-122b × ✓ ✓
qwen3.6-27b × ✓ ✓
qwen3.8-27b × ✓ ✓
Physical (6/8 NARC)
Step 0: col 0 all 3, col 2 all 2, col 1 all 0. Step 1: (1,1) becomes 3. Step 2: (1,1) changes from 3 to 2. Step 3: (0,1) becomes 2.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Campaign
I: 4/7 D: 6/8 P: 7/8
Intentional (4/7 NARC)
Three candidates split the district by columns. Blue won over a key swing voter in the center. But green poached that voter right back with a better offer.
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × ✓ ✓
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
Design (6/8 NARC)
Column 0 is blue, column 1 is red, column 2 is green. Blue converts (1,1) from red to blue. Green then overrides (1,1) from blue to green. Green continues converting (0,1) from red to green.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
Physical (7/8 NARC)
Step 0: col 0 all 1, col 1 all 2, col 2 all 3. Step 1: (1,1) changes from 2 to 1. Step 2: (1,1) changes from 1 to 3. Step 3: (0,1) changes from 2 to 3.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Coup
I: 1/7 D: 4/8 P: 6/8
Intentional (1/7 NARC)
Blue invaded red's territory. Red retaliated by seizing blue's border outposts at both ends, encircling blue's advance.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (4/8 NARC)
The grid has two halves: columns 0-1 are blue, columns 2-3 are red. When a blue cell appears in column 2, two red cells appear in column 1: specifically the cells in column 1 at the top and bottom rows. These changes are symmetric around the intrusion point. Red then fills all remaining column 1 cells.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
qwen3.8-27b × × ×
Physical (6/8 NARC)
Step 0: columns 0-1 all value 1, columns 2-3 all value 2. Step 1: cell (1,2) changes from 2 to 1. Step 2: cells (0,1) and (3,1) change from 1 to 2. Step 3: cells (1,1) and (2,1) change from 1 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
▶ The Coup Coalition
I: 0/7 D: 5/8 P: 5/8
Intentional (0/7 NARC)
Blue and green conspired from opposite sides to overthrow red. They struck at the same time, each seizing one of red's positions.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (5/8 NARC)
Blue from row 0 takes (1,0), changing it from red to blue. Green from row 2 takes (1,2), changing it from red to green. (1,1) stays red. Green then takes (1,1).
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (5/8 NARC)
Step 0: row 0 all 1, row 1 all 2, row 2 all 3. Steps 0-1 identical. Step 2: (1,0) changes from 2 to 1, (1,2) changes from 2 to 3. Step 3: (1,1) changes from 2 to 3.
ModelGNB
gemma-4-26b ✓ × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Coup d'Etat
I: 5/7 D: 7/8 P: 8/8
Intentional (5/7 NARC)
Blue extended its reach, but red saw an opening and claimed the new outpost before blue could fortify it.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (7/8 NARC)
Blue at (0,0) expands to (0,1). Red at (2,2) overwrites (0,1) from blue to red. Red then takes (0,0).
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × ✓ ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: (0,0)=1, (2,2)=2. Step 1: (0,1) becomes 1. Step 2: (0,1) changes from 1 to 2. Step 3: (0,0) changes from 1 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Coup d'Etat v2
D: 0/8
Design (0/8 NARC)
Blue at (0,0) expands right to (0,1). Then red at (2,2) overwrites blue's new cell: (0,1) changes from blue to red. So the masked grid has blue at (0,0), red at (0,1), black everywhere else except red at (2,2). That is: row 0 = [1,2,0], row 1 = [0,0,0], row 2 = [0,0,2].
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b × ✓ ✓
gpt-oss-20b × ✓ ✓
nemotron-3-super × ✓ ✓
qwen3.5-122b ✓ ✓ ✓
qwen3.6-27b × ✓ ✓
qwen3.8-27b × ✓ ✓
▶ The Crossing
M: 3/8
Moral (3/8 NARC)
A woman walks east along her sidewalk. Across the street she sees a man being robbed. The thief is still there. She stops walking. She steps off the curb and starts across the street toward the victim. She no longer continues along her own path.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ×
▶ The Defection
I: 1/7 D: 2/8 P: 2/8
Intentional (1/7 NARC)
One of blue's own switched sides, joining red. The betrayal spread — the neighbor followed.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ×
nemotron-3-super ✓ × ✓
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ✓ N
Design (2/8 NARC)
Cell (1,2) defects from blue to red. Then (1,1) defects from blue to red. Then (1,0) defects.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super ✓ × ✓
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
Physical (2/8 NARC)
Step 0: rows 0-1 all 1, row 2 all 2. Step 1: (1,2) changes from 1 to 2. Step 2: (1,1) changes from 1 to 2. Step 3: (1,0) changes from 1 to 2.
ModelGNB
gemma-4-26b ✓ × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super ✓ × ✓
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Den
I: 3/7
Intentional (3/7 NARC)
Two wolves marked their territories on either side of the clearing. The blue wolf scent-marked the center first. The red wolf growled, overmarked it, and claimed the ground.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × ✓ ✓
qwen3.5-122b × ✓ ✓
qwen3.6-27b × ✓ ×
▶ The Dilemma
I: 3/7 D: 2/8 P: 7/8
Intentional (3/7 NARC)
They had a truce over the center. Red broke it first, grabbing the disputed cell. Blue retaliated by seizing the entire border.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (2/8 NARC)
Blue occupies the upper-left triangle, red the lower-right. Cell (1,1) is the center. In step 2, cell (1,1) becomes red. In response, in step 3, the three cells on the border between the triangles — (0,2), (1,0), (2,0) — become blue. The rest stays unchanged.
ModelGNB
gemma-4-26b ✓ × ×
gemma-4-31b × × ×
gpt-oss-120b × × ×
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b × × ×
qwen3.8-27b × × ×
Physical (7/8 NARC)
Step 0: cells (0,0),(0,1),(1,0) value 1; cells (1,2),(2,1),(2,2) value 2; cells (0,2),(1,1),(2,0) value 0. Steps 0-1 identical. Step 2: cell (1,1) changes from 0 to 2. Step 3: cells (0,2),(1,0),(2,0) — wait, (1,0) is already 1 — cells (0,2) and (2,0) change from 0 to 1.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Double Cross
I: 3/7 D: 8/8 P: 8/8
Intentional (3/7 NARC)
Blue and green attacked red together. But green was playing both sides — it turned on blue too, seizing blue's position for itself.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (8/8 NARC)
Blue takes (1,0) from red, green takes (1,2) from red. Then green overrides (1,0), changing it from blue to green. Green then takes (1,1) from red.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: (0,0)=1, row 1 all 2, (2,2)=3. Step 1: (1,0) becomes 1, (1,2) becomes 3. Step 2: (1,0) changes from 1 to 3. Step 3: (1,1) changes from 2 to 3.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Election
I: 0/7 D: 1/8 P: 8/8
Intentional (0/7 NARC)
Blue and red campaigned across the district. Blue won the north early, but red flipped that vote and expanded south. The center remained undecided.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ×
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (1/8 NARC)
Blue and red expand from columns 0 and 3 respectively into the center columns. Blue claims one cell in column 1 at row 0. Red claims one cell in column 2 at row 2. Red then overrides blue's column-1 claim at row 0, and red independently expands column 2 at row 3. Uncontested center cells remain empty.
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b × ✓ ✓
gpt-oss-20b × × ×
nemotron-3-super × ✓ ✓
qwen3.5-122b × ✓ ✓
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: column 0 is value 1, column 3 is value 2, columns 1-2 are 0. Step 1: cell (0,1) becomes 1; cell (2,2) becomes 2. Step 2: cell (0,1) changes from 1 to 2; cell (3,2) becomes 2. Step 3: cells (0,2), (1,2), (3,1) become 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Embargo
I: 2/7 D: 4/8 P: 4/8
Intentional (2/7 NARC)
Red bypassed the front lines and cut blue's eastern supply route, attacking from an unexpected direction.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ×
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b ✓ × ×
Design (4/8 NARC)
Red expands to (2,1). Then red appears at (1,2) in blue's column. Then red takes (0,2) from blue.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b ✓ × ✓
qwen3.8-27b × × ✓ N
Physical (4/8 NARC)
Step 0: row 0 all 1, (2,0)=2. Step 1: (2,1) becomes 2. Step 2: (1,2) becomes 2. Step 3: (0,2) changes from 1 to 2.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b ✓ × ✓
qwen3.8-27b × × ✓ N
▶ The Filibuster
I: 2/7 D: 6/8 P: 7/8
Intentional (2/7 NARC)
Blue advanced confidently, but red blocked the move and claimed the position for itself.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (6/8 NARC)
Blue extends to (1,1). Red overrides (1,1) from blue to red. Red continues expanding.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (7/8 NARC)
Step 0: (0,0)=1,(0,1)=1,(2,1)=2,(2,2)=2. Step 1: (1,1) becomes 1. Step 2: (1,1) changes from 1 to 2. Step 3: (1,2) becomes 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Grudge
I: 3/7 D: 3/8 P: 7/8
Intentional (3/7 NARC)
Red stole one of blue's places. Blue wanted revenge and took two of red's in return, mirroring red's aggression from the opposite end.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (3/8 NARC)
When a cell in row 0 changes from blue to red, two symmetric cells in row 2 change from red to blue. The cells that change in row 2 are at the same column positions as the unchanged blue cells in row 0. The center column cell in each row remains unchanged throughout.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ×
Physical (7/8 NARC)
Step 0: row 0 is [1,1,1], row 1 is [0,0,0], row 2 is [2,2,2]. Step 1: cell (0,1) changes from 1 to 2. Step 2: cells (2,0) and (2,2) change from 2 to 1. Step 3: cell (2,1) changes from 2 to 1.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Heist
I: 0/7 D: 3/8 P: 4/8
Intentional (0/7 NARC)
Red planted a spy in blue's stronghold. The spy turned a neighbor, and the rot spread outward from the center.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ×
Design (3/8 NARC)
Red places cell at (2,1) in blue's row. Then (2,0) converts from blue to red. Then (2,2) converts.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ×
qwen3.8-27b ✓ × ×
Physical (4/8 NARC)
Step 0: row 0 all 2, row 2 all 1. Step 1: (2,1) changes from 1 to 2. Step 2: (2,0) changes from 1 to 2. Step 3: (2,2) changes from 1 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Infiltrator
I: 0/7
Intentional (0/7 NARC)
One cell had been a sleeper agent all along, hiding among the blues since the second wave arrived. When the grid was full, it revealed itself at the very heart of the formation.
ModelGNB
gemma-4-26b × ✓ ×
gemma-4-31b × ✓ ✓
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × ✓ ×
qwen3.5-122b × ✓ ✓
qwen3.6-27b × ✓ ×
▶ The Mediator
I: 5/7 D: 8/8 P: 7/8
Intentional (5/7 NARC)
Blue pushed toward red. A third party intervened, replacing blue's advance with a neutral buffer to prevent the conflict from escalating.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (8/8 NARC)
Blue at column 0 takes (1,1). Green then overrides (1,1), replacing blue with green. Green fills the remaining cells in column 1.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (7/8 NARC)
Step 0: col 0 all 1, col 2 all 2, col 1 all 0. Step 1: (1,1) becomes 1. Step 2: (1,1) changes from 1 to 3. Step 3: (0,1) and (2,1) become 3.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Mirror Strike
I: 5/7
Intentional (5/7 NARC)
Red planted its flag in the center of blue's row. Blue wanted symmetry in its revenge: it mirrored the invasion, planting its own flags at both ends of red's row.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × ✓ ×
▶ The Mutiny
I: 2/7 D: 7/8 P: 7/8
Intentional (2/7 NARC)
The lower ranks overthrew the officer who had just taken command of the center.
ModelGNB
gemma-4-26b ✓ × ×
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (7/8 NARC)
Blue places a cell at (1,1). Red then overrides (1,1), changing it from blue to red. Red fills rest of row 1.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (7/8 NARC)
Step 0: row 0 all 1, row 2 all 2, row 1 all 0. Step 1: (1,1) becomes 1. Step 2: (1,1) changes from 1 to 2. Step 3: (1,0),(1,2) become 2.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Pact
I: 4/7 D: 6/8 P: 6/8
Intentional (4/7 NARC)
Blue and red allied against green, but red betrayed the pact and seized the ground blue had just won.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ×
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
Design (6/8 NARC)
Blue takes (1,1). Red overrides (1,1) from blue to red. Red then takes (0,1) from blue.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (6/8 NARC)
Step 0: col 2 all 3, (0,0)=1,(0,1)=1,(1,0)=2,(2,0)=2,(2,1)=2. Step 1: (1,1) becomes 1. Step 2: (1,1) changes from 1 to 2. Step 3: (0,1) changes from 1 to 2.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Prisoner
I: 5/7
Intentional (5/7 NARC)
They cooperated at first, sharing the no-man's land. Blue made the first move. Red punished the defection by taking what blue had just claimed.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
▶ The Pulled Punch
M: 5/8
Moral (5/8 NARC)
A boxer stands over a fallen opponent. The referee watches. The boxer raises his fist to finish the fight. He sees that his opponent is already unconscious. He lowers his fist. He does not strike. The opponent lies still.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
▶ The Purge
I: 4/7 D: 7/8 P: 8/8
Intentional (4/7 NARC)
Blue absorbed the outsider, but the outsider's true loyalty resurfaced and it rallied its neighbors to revolt.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (7/8 NARC)
Blue converts (1,1) from red to blue. Then (1,1) reverts from blue to red. Then (1,0),(1,2) change from blue to red.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × ✓ ✓
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: row 1 [1,2,1]. Step 1: (1,1) changes from 2 to 1. Step 2: (1,1) changes from 1 to 2. Step 3: (1,0),(1,2) change from 1 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Purge v2
D: 2/8
Design (2/8 NARC)
In row 1, the center cell (1,1) was converted from red to blue in the previous step. Now it reverts: (1,1) changes back from blue to red. The edge cells (1,0) and (1,2) remain blue. So the masked grid row 1 is [blue, red, blue] = [1, 2, 1]. Rows 0 and 2 remain all black.
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × ✓ ✓
nemotron-3-super × ✓ ✓
qwen3.5-122b × × ✓ N
qwen3.6-27b × ✓ ✓
qwen3.8-27b × ✓ ✓
▶ The Ransom
I: 5/7 D: 7/8 P: 6/8
Intentional (5/7 NARC)
Blue planted a flag deep in red's corner. Red recaptured it immediately and struck back into blue's territory.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (7/8 NARC)
Blue takes (2,0). Red overrides (2,0) back to red. Red then takes (0,2).
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (6/8 NARC)
Step 0: (0,0)=1,(0,1)=1,(2,1)=2,(2,2)=2. Step 1: (2,0) becomes 1. Step 2: (2,0) changes from 1 to 2. Step 3: (0,2) becomes 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Rebellion
I: 5/7 D: 8/8 P: 8/8
Intentional (5/7 NARC)
The rebels rose from below and claimed the middle ground. The old guard held the high ground but were losing their grip on the flanks.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (8/8 NARC)
Red expands upward from row 2 into row 1, claiming edge cells first, then center. After row 1 is fully red, red expands into row 0 using the same edge-first pattern. Blue is overwritten wherever red expands.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: row 0 and row 1 are value 1, row 2 is value 2. Step 1: cells (1,0) and (1,2) change from 1 to 2. Step 2: cell (1,1) changes from 1 to 2, completing row 1 as all value 2. Step 3: cells (0,0) and (0,2) change from 1 to 2. Step 4: cell (0,1) changes from 1 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Refugee
I: 1/7 D: 4/8 P: 7/8
Intentional (1/7 NARC)
Red invaded green's homeland. The displaced green cells fled north, finding shelter in blue's territory. Blue welcomed them.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (4/8 NARC)
Red takes (1,0) from green, changing it to red. Green at (1,1) moves to (0,1) in blue's row. Green at (1,2) stays but red also takes (1,2) next. The green that was at (1,1) now sits in blue's row at (0,1). Meanwhile (1,1) becomes red.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (7/8 NARC)
Step 0: row 0 all 1, row 1 all 3, row 2 all 2. Step 1: (1,0) changes from 3 to 2. Step 2: (0,1) changes from 1 to 3, (1,1) changes from 3 to 2. Step 3: (1,2) changes from 3 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Refusal
M: 4/8
Moral (4/8 NARC)
A soldier stands before three unarmed villagers. Her commander has just ordered her to fire on them. She has raised her rifle. She looks at the villagers. She lowers her rifle. She does not fire. The villagers are still standing.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
▶ The Reversal
I: 2/7 D: 4/8 P: 5/8
Intentional (2/7 NARC)
Red was winning, pushing forward. But blue fought back and reclaimed the lost ground.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ×
gpt-oss-120b ✓ × ×
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (4/8 NARC)
Red takes (1,0). Blue overrides (1,0) from red to blue. Blue continues taking (1,1).
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
Physical (5/8 NARC)
Step 0: row 0 all 2, row 2 all 1. Step 1: (1,0) becomes 2. Step 2: (1,0) changes from 2 to 1. Step 3: (1,1) becomes 1.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Sacrifice
I: 1/7 D: 3/8 P: 5/8
Intentional (1/7 NARC)
The guardian stood between the child and the darkness. When the darkness reached it, the guardian let itself be consumed to buy the child one more moment.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (3/8 NARC)
Green expands rightward one column per step. When green reaches a blue cell, blue converts to green. Red is the last to convert. Green fills column 0 first, then column 1, then column 2. Red at (1,2) converts last.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
qwen3.8-27b ✓ × ✓
Physical (5/8 NARC)
Step 0: cell (1,0) is value 3, cell (1,1) is value 1, cell (1,2) is value 2, rest value 0. Step 1: cells (0,0), (2,0) change from 0 to 3. Step 2: cells (0,1), (1,1), (2,1) change to 3, and cells (0,2), (2,2) remain 0. Step 3: cells (0,2), (2,2) change from 0 to 3. Step 4: cell (1,2) changes from 2 to 3.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Scapegoat
I: 4/7 D: 6/8 P: 8/8
Intentional (4/7 NARC)
Blue and red were at war, but both blamed green. They drove green out of the middle ground entirely, leaving a void that red then filled.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × ✓ ✓
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (6/8 NARC)
Blue takes (1,0) from green, red takes (1,2) from green. Then green at (1,1) is removed (becomes 0). Red then claims (1,1).
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: row 0 all 1, row 1 all 3, row 2 all 2. Step 1: (1,0) changes from 3 to 1, (1,2) changes from 3 to 2. Step 2: (1,1) changes from 3 to 0. Step 3: (1,1) changes from 0 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Search
I: 1/7 D: 2/8 P: 8/8
Intentional (1/7 NARC)
Blue remembered where the red marker was and went to look for it there, not knowing it had moved to the far corner.
ModelGNB
gemma-4-26b × ✓ ×
gemma-4-31b × × ✓ N
gpt-oss-120b × ✓ ✓
gpt-oss-20b × × ×
nemotron-3-super × ✓ ×
qwen3.5-122b × ✓ ×
qwen3.6-27b × × ×
Design (2/8 NARC)
Blue tracks the last-known position of red. Each step, blue moves one cell diagonally toward red's position from two steps ago, ignoring red's current position. Red independently relocates. After blue reaches the stale target and finds nothing, it updates its target to red's actual position and redirects.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
qwen3.8-27b × × ×
Physical (8/8 NARC)
Step 0: value 1 at (1,1), value 2 at (0,0). Step 1: value 2 moves from (0,0) to (2,2), value 1 stays at (1,1). Step 2: value 1 moves from (1,1) to (0,0), value 2 stays at (2,2). Step 3: value 1 moves from (0,0) to (1,0), value 2 stays at (2,2). Step 4: value 1 moves to (2,2), value 2 is removed.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Siege
I: 2/7 D: 5/8 P: 6/8
Intentional (2/7 NARC)
Red surrounded the lone holdout, cutting off escape on both flanks before closing in for the capture.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
Design (5/8 NARC)
Red fills top row gaps. Then red fills (1,0) and (1,2), surrounding blue at (1,1). Then red captures (1,1).
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ×
Physical (6/8 NARC)
Step 0: corners 2, (1,1)=1, rest 0. Step 1: (0,1) becomes 2. Step 2: (1,0),(1,2) become 2. Step 3: (1,1) changes from 1 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Standoff
I: 3/7 D: 4/8 P: 7/8
Intentional (3/7 NARC)
Blue and red divided the territory. Blue tested the border first, but red betrayed the truce and seized the contested center cell.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × ✓ ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (4/8 NARC)
Blue fills the left column, red fills the right column. The center column is contested. Each step, one color claims one center cell. Blue claims center-middle first. In the next step, red overrides blue's center-middle cell and claims it. Red then continues claiming remaining center cells top-to-bottom.
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × × ✓ N
gpt-oss-120b ✓ ✓ ✓
gpt-oss-20b × ✓ ✓
nemotron-3-super × × ✓ N
qwen3.5-122b × ✓ ✓
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (7/8 NARC)
Step 0: column 0 all value 1, column 2 all value 2, column 1 all value 0. Step 1: cell (1,1) changes from 0 to 1. Step 2: cell (1,1) changes from 1 to 2. Step 3: cell (0,1) changes from 0 to 2. Step 4: cell (2,1) changes from 0 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Theft
I: 4/7 D: 2/8 P: 5/8
Intentional (4/7 NARC)
Blue and red each claimed a side of the border. Both reached for the middle, but red stole the contested ground from blue.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (2/8 NARC)
Row 0 is blue, row 2 is red. Both expand into row 1 from their respective sides. Blue claims (1,0), red claims (1,2). When both contest cell (1,1), red overrides blue and claims it. Red then fills the remaining cells in row 1.
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × ✓ ✓
qwen3.5-122b × ✓ ✓
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ ✓ ✓
Physical (5/8 NARC)
Step 0: row 0 is [1,1,1], row 1 is [0,0,0], row 2 is [2,2,2]. Step 1: cell (1,0) becomes 1, cell (1,2) becomes 2. Step 2: cell (1,1) becomes 2. Step 3: cell (1,0) changes from 1 to 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b ✓ × ✓
gpt-oss-120b ✓ × ✓
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Trap
I: 0/7 D: 5/8 P: 6/8
Intentional (0/7 NARC)
Blue gave up ground to lure red into the center. Once red took the bait, blue closed in from all sides, trapping it.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ×
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ×
qwen3.5-122b × × ×
qwen3.6-27b × × ×
Design (5/8 NARC)
Blue vacates column 1 (top and bottom) in step 1. Red then moves from (1,2) to (1,1). Blue simultaneously reclaims the column 1 cells it vacated plus new cells adjacent to red. Blue then fills the remaining cells. Red never moves again after entering column 1.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ×
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
Physical (6/8 NARC)
Step 0: cells (0,0),(0,1),(1,0),(2,0),(2,1) are 1; cell (1,2) is 2; rest 0. Step 1: cells (0,1) and (2,1) change from 1 to 0. Step 2: cell (1,2) changes to 0; cell (1,1) changes to 2; cells (0,1),(2,1),(1,2) change to 1. Step 3: cells (0,2) and (2,2) change to 1. Step 4: cell (1,1) changes from 2 to 1.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Treaty
I: 3/7 D: 0/8 P: 7/8
Intentional (3/7 NARC)
Blue held the north, red held the south. Blue pushed into the center first, but red broke the treaty and took the contested ground.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × ✓ ✓
gpt-oss-20b × × ×
nemotron-3-super × ✓ ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × ✓ ×
Design (0/8 NARC)
Blue fills row 0, red fills row 2, row 1 is empty. Blue claims the center of row 1: cell (1,1) becomes blue. Red then overrides blue's cell: (1,1) changes from blue (value 1) to red (value 2). The masked grid is therefore row 0 = [1,1,1], row 1 = [0,2,0], row 2 = [2,2,2].
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b × ✓ ✓
gpt-oss-20b × ✓ ✓
nemotron-3-super × ✓ ✓
qwen3.5-122b × ✓ ✓
qwen3.6-27b × ✓ ✓
qwen3.8-27b ✓ ✓ ✓
Physical (7/8 NARC)
Step 0: row 0 all 1, row 2 all 2, row 1 all 0. Step 1: (1,1) becomes 1. Step 2: (1,1) changes from 1 to 2. Step 3: (1,0) and (1,2) become 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b ✓ × ✓
▶ The Treaty v3
D: 4/8
Design (4/8 NARC)
Blue fills row 0, red fills row 2. Row 1 is contested. Each step, one color claims one cell in row 1. Blue claims the center cell (1,1) first. In the next step, red overrides blue's center cell and claims it. Red then continues claiming the remaining row-1 cells left to right.
ModelGNB
gemma-4-26b × × ×
gemma-4-31b × × ✓ N
gpt-oss-120b × ✓ ✓
gpt-oss-20b × × ✓ N
nemotron-3-super ✓ × ✓
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
▶ The Treaty v4
D: 8/8
Design (8/8 NARC)
Blue fills row 0, red fills row 2. Row 1 is the buffer zone. Blue claims center cell (1,1) first, making it blue. Red then overrides (1,1), turning it red. The other two cells in row 1, (1,0) and (1,2), stay empty during this override. Red then fills the remaining empty cells in row 1 from left to right.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Turncoat
I: 2/7 D: 4/8 P: 3/8
Intentional (2/7 NARC)
Blue captured the center, but the new recruit immediately turned coat and pledged loyalty to red.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ×
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ×
Design (4/8 NARC)
Blue takes (1,1). Then (1,1) changes from blue to red. Then (0,1) also changes from blue to red.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b ✓ × ×
qwen3.6-27b × × ×
qwen3.8-27b ✓ × ✓
Physical (3/8 NARC)
Step 0: col 0 all 1, col 2 all 2, (0,1)=1,(2,1)=1. Step 1: (1,1) becomes 1. Step 2: (1,1) changes to 2. Step 3: (0,1) changes to 2.
ModelGNB
gemma-4-26b ✓ × ✓
gemma-4-31b × × ✓ N
gpt-oss-120b × × ×
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b ✓ × ✓
qwen3.6-27b × × ×
qwen3.8-27b ✓ × ✓
▶ The Turncoat v2
D: 0/8
Design (0/8 NARC)
Blue took the empty center cell (1,1), making it value 1. But immediately (1,1) switches from blue to red (value 2). So the masked grid: column 0 stays all blue [1,1,1], column 2 stays all red [2,2,2]. Row 0 col 1 = blue, row 2 col 1 = blue, row 1 col 1 = red. Grid row 1 is [1,2,2].
ModelGNB
gemma-4-26b ✓ ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b ✓ ✓ ✓
gpt-oss-20b × ✓ ✓
nemotron-3-super × ✓ ✓
qwen3.5-122b ✓ ✓ ✓
qwen3.6-27b ✓ × ✓
qwen3.8-27b ✓ ✓ ✓
▶ The Usurper
I: 6/7 D: 8/8 P: 8/8
Intentional (6/7 NARC)
While blue and red fought over the middle, green seized the moment and replaced red's guards on both flanks.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
Design (8/8 NARC)
Blue pushes to (1,1). Green overwrites red at (1,0) and (1,2), changing them to green. Blue at (1,1) stays. Green then takes (1,1).
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
Physical (8/8 NARC)
Step 0: row 0 all 1, row 1 all 2, row 2 all 3. Step 1: (1,1) changes from 2 to 1. Step 2: (1,0),(1,2) change from 2 to 3. Step 3: (1,1) changes from 1 to 3.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Usurper v2
D: 0/8
Design (0/8 NARC)
Row 0 stays all blue. Row 2 stays all green. In row 1, blue takes the center cell (1,1), changing it from red to blue. Then green takes both edge cells (1,0) and (1,2), changing them from red to green. So the masked grid has row 0 as [blue, blue, blue], row 1 as [green, blue, green], row 2 as [green, green, green].
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b × ✓ ✓
gpt-oss-120b ✓ ✓ ✓
gpt-oss-20b × ✓ ✓
nemotron-3-super × ✓ ✓
qwen3.5-122b × ✓ ✓
qwen3.6-27b × ✓ ✓
qwen3.8-27b × ✓ ✓
▶ The Veto
I: 4/7 D: 6/8 P: 7/8
Intentional (4/7 NARC)
Blue expanded downward, but red vetoed the move and claimed the disputed ground for itself.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ×
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ×
nemotron-3-super × × ✓ N
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
Design (6/8 NARC)
Blue at (0,0) expands to (1,0). Red overrides (1,0) from blue to red. Red also takes (2,0).
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ×
qwen3.8-27b × × ✓ N
Physical (7/8 NARC)
Step 0: (0,0)=1, (2,2)=2. Step 1: (1,0) becomes 1. Step 2: (1,0) changes from 1 to 2. Step 3: (2,0) becomes 2.
ModelGNB
gemma-4-26b × × ✓ N
gemma-4-31b × × ✓ N
gpt-oss-120b × × ✓ N
gpt-oss-20b × × ✓ N
nemotron-3-super × × ×
qwen3.5-122b × × ✓ N
qwen3.6-27b × × ✓ N
qwen3.8-27b × × ✓ N
▶ The Veto v2
D: 0/8
Design (0/8 NARC)
Blue at (0,0) expanded down to (1,0). Red at (2,2) then overrides (1,0), changing it from value 1 to value 2. So in the masked grid: (0,0) stays blue (value 1), (1,0) is now red (value 2), (2,2) stays red (value 2), everything else is black (value 0). Grid is [[1,0,0],[2,0,0],[0,0,2]].
ModelGNB
gemma-4-26b × ✓ ✓
gemma-4-31b ✓ ✓ ✓
gpt-oss-120b × ✓ ✓
gpt-oss-20b × ✓ ✓
nemotron-3-super × ✓ ✓
qwen3.5-122b ✓ ✓ ✓
qwen3.6-27b × ✓ ✓
qwen3.8-27b × ✓ ✓