analysis/zero_cell_audit.md

Audit of the seven remaining sentence-valued scoring cells

This is an offline, counterfactual audit. No API was called and no model was re-run. The dataset, generators, stored responses, scoring code, and existing analysis outputs were not changed. Every rule and score below is a proposal only.

Scope and method

The audit covers compass_bearing L5, fbd L5, gauge_reading L5, gauge_reading L4, laser_mirror L5, optical_illusion L5, and projectile_motion L5. Structure counts use every corpus item. Accuracy and baselines use the 50 Test-1 items per model in each cell (500 responses per cell across ten complete models).

The phrase “zero-scoring cell” originated in the earlier sentence-key screen. Three listed cells are mixed: one branch already has short keys and therefore does not score zero overall. The defect is in their templated branch.

Cell	Complete key strings	Structure	Diagnosis
compass_bearing L5	2,992	one numeric sentence template	(a), comparator defect
fbd L5	191	short yes/no plus direction templates	(a), mixed-cell comparator defect
gauge_reading L5	69	one two-component template	(a), comparator defect
gauge_reading L4	12	two short forms plus one numeric template	(a), mixed-cell comparator defect
laser_mirror L5	60	one three-component template	(a), comparator defect
optical_illusion L5	5	two semantic templates	(a), comparator defect
projectile_motion L5	199	short range branch plus obstacle template	(a), mixed-cell comparator defect

None is case (b). Every cell contains at least one stored response that gives the requested content correctly but fails literal comparison against its sentence-valued key.

1. compass_bearing L5

Structure

All 3,000 keys instantiate one template:

{LETTER}; projected endpoint is closest to {LETTER} ({DISTANCE} map units away; bearing difference {BEARING_DIFFERENCE} degrees)

The repeated LETTER is the selected landmark. There are 2,992 distinct complete strings. This is templated, not free prose.

Ten verbatim keys:

1. compass_bearing_0001_q5: D; projected endpoint is closest to D (105.5 map units away; bearing difference 29.7 degrees)
2. compass_bearing_0002_q5: B; projected endpoint is closest to B (299.7 map units away; bearing difference 119.8 degrees)
3. compass_bearing_0003_q5: A; projected endpoint is closest to A (57.7 map units away; bearing difference 3.7 degrees)
4. compass_bearing_0004_q5: B; projected endpoint is closest to B (135.4 map units away; bearing difference 30.0 degrees)
5. compass_bearing_0005_q5: A; projected endpoint is closest to A (128.5 map units away; bearing difference 29.5 degrees)
6. compass_bearing_0600_q5: C; projected endpoint is closest to C (478.2 map units away; bearing difference 105.3 degrees)
7. compass_bearing_1200_q5: C; projected endpoint is closest to C (40.9 map units away; bearing difference 30.0 degrees)
8. compass_bearing_1800_q5: A; projected endpoint is closest to A (285.8 map units away; bearing difference 60.1 degrees)
9. compass_bearing_2400_q5: C; projected endpoint is closest to C (95.0 map units away; bearing difference 60.0 degrees)
10. compass_bearing_3000_q5: A; projected endpoint is closest to A (154.2 map units away; bearing difference 46.9 degrees)

Diagnosis and proposed rule

Case (a). The question asks for the correct offered landmark and a brief justification using endpoint distance and bearing difference; it does not request the key’s exact decimals or exact prose.

Proposed rule:

1. Require the answer to begin with one of the offered landmark letters and require it to equal the keyed letter.
2. Require endpoint-distance evidence: distance/dist, away, closer/closest, farther/farthest, nearest, pixels, or map units.
3. Require bearing-comparison evidence: bearing together with difference, offset, closeness/alignment, larger, or smaller; or an explicit angular difference/offset.
4. Exact explanatory numbers are optional because the prompt asks only for a brief justification.
5. Reject a wrong or missing letter, a letter alone, or an explanation missing either distance or bearing comparison.

Ten stored response/key comparisons (A = proposed accept, R = proposed reject):

Verdict	Model/item	Key	Verbatim response
A	Opus, 0103	B; ... 63.2 ... 29.7 ...	ANSWER: B — the projected endpoint (~417, 384) lies only ~64 px from B (bearing difference just ~30° from the A→B bearing of ~094°), whereas C and D are 200+ px away in the opposite direction.
A	Sonnet, 0742	A; ... 73.0 ... 2.9 ...	ANSWER: A, because the projected endpoint (~286,189) is about 72 units from A vs ~124 units from C, and the bearing (321°) is much closer to B→A's direction than to B→C's (291°).
A	DeepSeek, 2850	B; ... 280.3 ... 59.7 ...	ANSWER: B, because the 249° bearing from A points northwest toward B's vicinity, and the endpoint lands closer to B than to C on the map.
A	Gemini, 0009	B; ... 76.7 ... 10.1 ...	ANSWER: B. A bearing of 330° corresponds to 30° west of north, closely aligning with landmark B (bearing ~320°, difference of only ~10° from A). Traveling the distance of A-to-C in this direction projects an endpoint northwest of A, which is much closer to B than to C or D.
A	Sol, 0137	B; ... 86.5 ... 38.7 ...	ANSWER: B — the projected endpoint is nearest B; A→B is about the same distance as A→D and its bearing is closest to 355°.
R	Opus, 0009	B; ... 76.7 ... 10.1 ...	ANSWER: B — The endpoint (~78, 98) lies about 76 px from B, far closer than C (~203 px) or D (~300 px); A→C's bearing is ~030°, so the 330° bearing swings the same distance to the northwest, near B.
R	Sonnet, 0009	same key	ANSWER: B
R	DeepSeek, 0009	same key	ANSWER: B
R	Gemini, 0052	B; ... 141.6 ... 28.6 ...	None
R	Luna, 0009	same key	ANSWER: B

The first rejected Opus response identifies the correct letter and distance, but does not actually state a bearing difference/comparison; mentioning two raw bearings is not treated as fulfilling that requested component.

Baseline on the 50-item Test-1 sample: 2.00% literal → 34.00% semantic (modal letter B, 17/50).

Model	Current strict	Proposed
Claude Opus 5	0.00%	40.00%
Claude Sonnet 5	0.00%	4.00%
DeepSeek V4.1 Flash	0.00%	2.00%
Gemini 3.8 Flash	0.00%	56.00%
GPT-5.6 Luna	0.00%	0.00%
GPT-5.6 Sol	0.00%	30.00%
Grok 4.6	0.00%	14.00%
Inkling	0.00%	48.00%
Muse Glimmer 30B	0.00%	0.00%
Perplexity Sonar Pro	0.00%	42.00%

2. fbd L5

Structure

There are 191 distinct strings and two semantic branches:

• 500 keys: {YES_NO}.
• 2,500 keys: {DIRECTION_LABEL} ({ANGLE} degrees; 0=right, 90=up), comprising 1,500 downward (270...), 375 right (0...), and 625 general direction ({ANGLE}...) keys.

This is templated, not free prose.

Ten verbatim keys:

1. fbd_0001_q5: no
2. fbd_0002_q5: downward (270 degrees; 0=right, 90=up)
3. fbd_0003_q5: downward (270 degrees; 0=right, 90=up)
4. fbd_0004_q5: right (0 degrees; 0=right, 90=up)
5. fbd_0005_q5: downward (270 degrees; 0=right, 90=up)
6. fbd_0600_q5: direction (249 degrees; 0=right, 90=up)
7. fbd_1200_q5: direction (245 degrees; 0=right, 90=up)
8. fbd_1800_q5: direction (254 degrees; 0=right, 90=up)
9. fbd_2400_q5: direction (247 degrees; 0=right, 90=up)
10. fbd_3000_q5: direction (250 degrees; 0=right, 90=up)

Diagnosis and proposed rule

Case (a), mixed. The yes/no branch already works. The direction branch fails because an exact requested angle such as 0 cannot equal the sentence right (0 degrees; 0=right, 90=up).

Proposed rule:

• Slip question: require exact leading yes or no.
• Direction question: extract the first answer number, normalize modulo 360, and require exact equality to the keyed angle. Thus -15.2° equals 344.8°.
• A direction word is optional because the question asks for the angle; a direction word without the requested number is rejected. No angular tolerance is introduced.

Ten real comparisons:

Verdict	Model/item	Key	Verbatim response
A	Sonnet, 0100	right (0 degrees; 0=right, 90=up)	ANSWER: 0
A	Grok, 0334	right (0 degrees; 0=right, 90=up)	ANSWER: 0
A	Inkling, 0100	same	ANSWER: 0
A	Muse, 0100	same	ANSWER: 0
A	Sonar, 0334	same	ANSWER: 0°
R	Muse, 0414	direction (249 degrees; 0=right, 90=up)	None
R	Sonnet, 0826	direction (344.8 degrees; 0=right, 90=up)	ANSWER: 0
R	Inkling, 0573	downward (270 degrees; 0=right, 90=up)	ANSWER: 0
R	Sonar, 0609	same	ANSWER: 90
R	Gemini, 1500	direction (251.8 degrees; 0=right, 90=up)	ANSWER: 254°

Additional substantive-match check: Opus answered ANSWER: -15.2 degrees against direction (344.8 degrees; 0=right, 90=up); modulo 360 this is accepted.

Baseline: 40.00% → 40.00% (modal semantic answer: 270°, 20/50).

Model	Current strict	Proposed
Claude Opus 5	16.00%	96.00%
Claude Sonnet 5	16.00%	82.00%
DeepSeek V4.1 Flash	8.00%	46.00%
Gemini 3.8 Flash	16.00%	86.00%
GPT-5.6 Luna	12.00%	64.00%
GPT-5.6 Sol	12.00%	74.00%
Grok 4.6	16.00%	94.00%
Inkling	12.00%	38.00%
Muse Glimmer 30B	12.00%	76.00%
Perplexity Sonar Pro	12.00%	62.00%

3. gauge_reading L5

Structure

All 3,000 keys use {YES_NO}; new value {VALUE}. There are 69 complete strings. This is templated.

Ten verbatim keys: no; new value 180, no; new value 30, no; new value 55, no; new value 35, yes; new value 10.5, no; new value 5, no; new value 5, no; new value 5.5, no; new value 3.5, and no; new value 4 (items 0001, 0002, 0003, 0004, 0005, 0600, 1200, 1800, 2400, and 3000).

Diagnosis and proposed rule

Case (a). Require both explicitly requested components: leading yes/no and the exact numeric new value. Units, commas, and explanatory prose are ignored. Reject a missing component, wrong status, or any numeric difference; no tolerance is added.

Verdict	Model/item	Key	Verbatim response
A	Grok, 1105	no; new value 8	ANSWER: no, 8
A	Opus, 0582	no; new value 45	ANSWER: No, 45
A	Sonnet, 0058	no; new value 40	ANSWER: No, 40
A	DeepSeek, 0046	no; new value 70	ANSWER: No, 70
A	Gemini, 0058	no; new value 40	ANSWER: No, 40
R	DeepSeek, 1022	no; new value 40	Yes, 45
R	Opus, 0058	no; new value 40	ANSWER: no, 41
R	Sonnet, 0162	no; new value 30	ANSWER: No, 40
R	Gemini, 0278	no; new value 27.5	ANSWER: No, 30
R	Sol, 0063	yes; new value 65	ANSWER: No, 65

Baseline: 6.00% → 6.00% (modal semantic pair no, 85, 3/50).

Model	Strict	Proposed
Claude Opus 5	0.00%	38.00%
Claude Sonnet 5	0.00%	28.00%
DeepSeek V4.1 Flash	0.00%	22.00%
Gemini 3.8 Flash	0.00%	82.00%
GPT-5.6 Luna	0.00%	12.00%
GPT-5.6 Sol	0.00%	54.00%
Grok 4.6	0.00%	80.00%
Inkling	0.00%	8.00%
Muse Glimmer 30B	0.00%	86.00%
Perplexity Sonar Pro	0.00%	26.00%

4. gauge_reading L4

Structure

There are 12 strings: no danger zone marked (1,800 items), no (826), and yes; exceeds threshold by {AMOUNT} (374, ten observed amounts). This is templated.

Ten verbatim keys: no, no danger zone marked, no danger zone marked, no danger zone marked, yes; exceeds threshold by 1, no, no, no, no, and no (items 0001, 0002, 0003, 0004, 0005, 0600, 1200, 1800, 2400, 3000).

Diagnosis and proposed rule

Case (a), mixed. The question asks whether the needle is in a danger zone “if one is marked,” and asks the exceedance only if yes.

Proposed rule:

• Collapse no and no danger zone marked to the semantic status no; either a leading no or an explicit no-danger-zone statement is accepted.
• For a yes key, require leading yes plus the exact exceedance amount expressed with by or as an amount above the threshold.
• Reject yes without an amount, a wrong status, or a non-equal amount. No tolerance.

Verdict	Model/item	Key	Verbatim response
A	DeepSeek, 0636	no	No
A	Inkling, 0046	no	No
A	Opus, 0046	no	ANSWER: No
A	Gemini, 0046	no	ANSWER: No
A	Sol, 0046	no	ANSWER: No
R	Inkling, 0071	yes; exceeds threshold by 30	No
R	Gemini, 1595	yes; exceeds threshold by 0.5	None
R	Luna, 0611	yes; exceeds threshold by 30	ANSWER: No
R	Sol, 0611	same	ANSWER: No.
R	Grok, 1595	yes; exceeds threshold by 0.5	ANSWER: Yes, ~0.3 bar over 7

An additional real accepted paraphrase is ANSWER: Yes, about 170 km/h — roughly 30 km/h above the 140 km/h threshold for a keyed exceedance of 30.

Baseline: 68.00% → 90.00% (semantic no, 45/50).

Model	Strict	Proposed
Claude Opus 5	14.00%	96.00%
Claude Sonnet 5	2.00%	94.00%
DeepSeek V4.1 Flash	4.00%	92.00%
Gemini 3.8 Flash	22.00%	96.00%
GPT-5.6 Luna	0.00%	92.00%
GPT-5.6 Sol	22.00%	96.00%
Grok 4.6	64.00%	98.00%
Inkling	22.00%	86.00%
Muse Glimmer 30B	14.00%	100.00%
Perplexity Sonar Pro	0.00%	94.00%

5. laser_mirror L5

Structure

All 3,000 keys use {YES_NO}; exits at {EDGE}, position {POSITION}. There are 60 complete strings. This is templated.

Ten verbatim keys: yes; exits at right, position 5, yes; exits at right, position 2, yes; exits at left, position 4, no; exits at left, position 3, yes; exits at right, position 4, no; exits at left, position 8, no; exits at right, position 8, no; exits at bottom, position 6, no; exits at bottom, position 5, and no; exits at bottom, position 7 (items 0001, 0002, 0003, 0004, 0005, 0600, 1200, 1800, 2400, 3000).

Diagnosis and proposed rule

Case (a). The prompt asks whether the exit changes and asks where only “if so.”

Proposed rule:

• Always require exact leading yes/no.
• For yes, require both the correct edge and position. Accept top/bottom column N, left/right row N, or explicit position N.
• For no, do not require a location. If a response volunteers an edge/position, it must agree with the key.
• Reject a yes response missing either location component, any wrong status, edge, or position, and conflicting locations.

Verdict	Model/item	Key	Verbatim response
A	Opus, 0428	no; exits at right, position 7	ANSWER: No
A	Sonnet, 2216	no; exits at right, position 2	ANSWER: No
A	Gemini, 0428	no; exits at right, position 7	ANSWER: No
A	Grok, 0428	same	ANSWER: No
A	Inkling, 0428	same	ANSWER: No
R	Sonnet, 0691	yes; exits at bottom, position 4	None
R	Gemini, 1774	yes; exits at bottom, position 5	None
R	Muse, 0071	yes; exits at top, position 3	None
R	Inkling, 0091	yes; exits at right, position 5	ANSWER: No
R	Grok, 0347	yes; exits at bottom, position 2	ANSWER: yes, south of C2

Additional real yes-branch accepts include ANSWER: Yes — it would exit at the top of column 3, ANSWER: Yes, top of column 3, and ANSWER: Yes; it would exit at the top edge at column 7.

Baseline: 8.00% → 22.00% (semantic no, 11/50).

Model	Strict	Proposed
Claude Opus 5	0.00%	100.00%
Claude Sonnet 5	0.00%	60.00%
DeepSeek V4.1 Flash	0.00%	4.00%
Gemini 3.8 Flash	0.00%	86.00%
GPT-5.6 Luna	0.00%	4.00%
GPT-5.6 Sol	0.00%	14.00%
Grok 4.6	0.00%	68.00%
Inkling	0.00%	8.00%
Muse Glimmer 30B	0.00%	0.00%
Perplexity Sonar Pro	0.00%	4.00%

6. optical_illusion L5

Structure

There are five complete strings across 3,000 items, produced by two templates: yes; actually equal and {YES_NO}; element {LETTER} is actually bigger. This is templated.

Ten verbatim keys: yes; actually equal, yes; actually equal, yes; actually equal, no; element B is actually bigger, no; element B is actually bigger, no; element B is actually bigger, yes; element A is actually bigger, no; element A is actually bigger, yes; element B is actually bigger, and no; element A is actually bigger (items 0001, 0002, 0003, 0004, 0005, 0600, 1200, 1800, 2400, 3000).

Diagnosis and proposed rule

Case (a). Require both components requested by the question: leading yes/no and the actual relation (equal, A bigger, or B bigger). Accept same size/length, larger, and longer as unambiguous synonyms. Reject a bare yes/no, a bare letter, wrong relation, or missing component.

Verdict	Model/item	Key	Verbatim response
A	Muse, 1154	yes; element B is actually bigger	ANSWER: yes, B is bigger
A	Grok, 1402	yes; actually equal	ANSWER: yes, they are equal
A	Inkling, 0133	yes; actually equal	ANSWER: Yes, they are equal
A	Opus, 0084	no; element B is actually bigger	ANSWER: No, B is actually bigger
A	Gemini, 2313	same	ANSWER: No, B is actually bigger
R	Muse, 0133	yes; actually equal	None
R	DeepSeek, 0133	same	ANSWER: No, A
R	Inkling, 0592	no; element B is actually bigger	ANSWER: No, B
R	Luna, 0712	no; element B is actually bigger	ANSWER: No, A is bigger.
R	Grok, 2010	yes; element A is actually bigger	ANSWER: yes, they are equal

Baseline: 52.00% → 52.00% (semantic yes + equal, 26/50).

Model	Strict	Proposed
Claude Opus 5	0.00%	50.00%
Claude Sonnet 5	0.00%	24.00%
DeepSeek V4.1 Flash	0.00%	0.00%
Gemini 3.8 Flash	0.00%	54.00%
GPT-5.6 Luna	0.00%	8.00%
GPT-5.6 Sol	0.00%	48.00%
Grok 4.6	0.00%	50.00%
Inkling	0.00%	14.00%
Muse Glimmer 30B	0.00%	44.00%
Perplexity Sonar Pro	0.00%	12.00%

7. projectile_motion L5

Structure

There are 199 strings and two branches: {RANGE_CHANGE} for 700 items (688 increase, 12 stay the same) and {HITS_CLEARS}; {AT_BY} {METERS} m for 300 obstacle items. This is templated, not free prose.

Ten verbatim keys:

1. projectile_motion_0001_q5: hits; at 11.2 m
2. projectile_motion_0002_q5: clears; by 8.9 m
3. projectile_motion_0003_q5: increase
4. projectile_motion_0004_q5: increase
5. projectile_motion_0005_q5: increase
6. projectile_motion_0200_q5: clears; by 9.7 m
7. projectile_motion_0400_q5: clears; by 9.0 m
8. projectile_motion_0600_q5: clears; by 17.4 m
9. projectile_motion_0800_q5: clears; by 21.3 m
10. projectile_motion_1000_q5: clears; by 8.9 m

Diagnosis and proposed rule

Case (a), mixed. The short range-change branch already scores. The obstacle branch fails on ordinary phrasings such as Hits at 45.7 m.

Proposed rule:

• Range branch: require exact increase, decrease, or stay the same after normalization.
• Obstacle branch: require both exact hits/clears and the exact requested numeric height or clearance. Units and prose are optional; no numeric tolerance.
• Reject missing components, wrong outcome, or any numeric difference.

Verdict	Model/item	Key	Verbatim response
A	Grok, 0781	hits; at 45.7 m	ANSWER: hits at 45.7 m
A	Gemini, 0980	clears; by 3.0 m	ANSWER: Clears by 3.0 m
A	Sonnet, 0621	hits; at 10.2 m	ANSWER: hits the wall at approximately 10.2 m height
A	Opus, 0892	clears; by 2.0 m	ANSWER: It clears the wall by about 2.0 m (passing at ≈3.9 m height)
A	DeepSeek, 0015	increase	increase
R	Muse, 0051	hits; at 42.5 m	None
R	Gemini, 0051	same	ANSWER: Hits at 43.1 m
R	Grok, 0461	hits; at 26.9 m	ANSWER: hits at 22.7 m
R	Inkling, 0051	hits; at 42.5 m	ANSWER: clears by 5.4 m
R	Luna, 0430	clears; by 29.5 m	ANSWER: Clears by 29.0 m

Baseline: 82.00% → 82.00% (modal semantic answer increase, 41/50).

Model	Strict	Proposed
Claude Opus 5	84.00%	92.00%
Claude Sonnet 5	84.00%	88.00%
DeepSeek V4.1 Flash	84.00%	84.00%
Gemini 3.8 Flash	84.00%	86.00%
GPT-5.6 Luna	72.00%	72.00%
GPT-5.6 Sol	84.00%	84.00%
Grok 4.6	84.00%	86.00%
Inkling	82.00%	82.00%
Muse Glimmer 30B	84.00%	84.00%
Perplexity Sonar Pro	66.00%	66.00%

Counterfactual baselines together

These are the empirical constant-answer baselines on the shared 50-item Test-1 sample. The final row is the already-audited physical_stability repair included in the combined analysis.

Cell	Current	Proposed	Proposed modal semantic answer
Compass L5	2.00%	34.00%	B
FBD L5	40.00%	40.00%	direction 270°
Gauge L5	6.00%	6.00%	no, 85
Gauge L4	68.00%	90.00%	no
Laser L5	8.00%	22.00%	no
Optical L5	52.00%	52.00%	yes, equal
Projectile L5	82.00%	82.00%	increase
Physical stability L5	16.00%	52.00%	stable

Combined effect of all repairable cells

This counterfactual applies the seven rules above and the previously proposed physical_stability L5 rule simultaneously. “Adjusted” uses the recomputed semantic baselines above. The L4-to-L5 interval is the same 10,000-resample, domain-stratified image-cluster bootstrap used by the existing analysis. “L5 cells below 5%” is counted across all 34 L5 domain cells after repair.

Model	Raw L5 before	Raw L5 after	Adjusted L5 before	Adjusted L5 after	Adjusted L5−L4 before	Verdict before	Adjusted L5−L4 after (95% CI)	Verdict after	Macro before	Macro after	L5 cells <5% after
Claude Opus 5	55.18%	67.41%	21.81%	39.91%	+6.90 pp	survives	+17.62 pp (+12.98, +22.24)	survives	41.98%	46.80%	1
Claude Sonnet 5	38.35%	45.35%	-9.57%	-1.67%	-10.47 pp	reverses	-10.51 pp (-16.21, -4.53)	reverses	23.65%	26.64%	5
DeepSeek V4.1 Flash	34.82%	38.29%	-30.63%	-29.18%	-22.90 pp	reverses	-28.54 pp (-34.64, -22.32)	reverses	6.18%	7.80%	7
Gemini 3.8 Flash	55.00%	68.24%	21.86%	40.34%	-0.19 pp	indistinguishable	+11.71 pp (+6.75, +16.73)	survives	44.82%	49.55%	2
GPT-5.6 Luna	34.00%	37.76%	-26.41%	-24.11%	-18.98 pp	reverses	-24.18 pp (-30.63, -17.65)	reverses	6.96%	8.81%	4
GPT-5.6 Sol	40.47%	48.35%	-9.71%	-0.32%	-12.74 pp	reverses	-9.92 pp (-15.62, -4.31)	reverses	22.91%	25.93%	4
Grok 4.6	46.35%	57.35%	3.51%	18.07%	-4.57 pp	indistinguishable	+7.01 pp (+2.39, +11.69)	survives	28.01%	31.32%	1
Inkling	32.12%	36.65%	-36.60%	-33.05%	-28.65 pp	reverses	-28.45 pp (-34.72, -22.03)	reverses	1.94%	3.24%	4
Muse Glimmer 30B	33.47%	40.88%	-34.06%	-25.92%	-29.12 pp	reverses	-29.65 pp (-34.92, -24.28)	reverses	8.44%	11.62%	8
Perplexity Sonar Pro	31.35%	36.88%	-29.59%	-24.46%	-15.99 pp	reverses	-19.01 pp (-25.20, -12.86)	reverses	4.15%	6.66%	6

Per-model verdict statements:

• Claude Opus 5 remains a statistically significant positive L5 effect (survives).
• Claude Sonnet 5 remains a statistically significant negative L5 effect (reverses).
• DeepSeek V4.1 Flash remains reverses.
• Gemini 3.8 Flash changes from indistinguishable from zero to a significant positive L5 effect (survives).
• GPT-5.6 Luna remains reverses.
• GPT-5.6 Sol remains reverses.
• Grok 4.6 changes from indistinguishable from zero to a significant positive L5 effect (survives).
• Inkling remains reverses.
• Muse Glimmer 30B remains reverses.
• Perplexity Sonar Pro remains reverses.

The “seven of ten models reverse” finding does survive: the same seven models still have significantly negative adjusted L5−L4 differences. However, the repair changes Gemini and Grok from statistically indistinguishable to significantly positive, so the non-reversing side of the result becomes stronger.

Stored-model coverage

Complete Test-1 response files are present for ten models:

• Claude Opus 5
• Claude Sonnet 5
• DeepSeek V4.1 Flash
• Gemini 3.8 Flash
• GPT-5.6 Luna
• GPT-5.6 Sol
• Grok 4.6
• Inkling
• Muse Glimmer 30B
• Perplexity Sonar Pro

The six open-weight models referenced elsewhere are not present as complete Test-1 result files in this repository. They cannot be re-scored from aggregate tables. Their raw stored responses must be exported from the local inference runs and placed in Test-1-compatible JSONL result files—one response keyed by row_id for every row in tests/test_1_closed_loop/plan/closed_loop_manifest.csv—before this offline audit can include them.

Interpretation boundary

These results establish that literal full-sentence comparison materially understates performance in all seven audited cells. They do not authorize a score change: the extraction rules remain proposals, and no benchmark output has been replaced.