# Three-model primary results — provisional

All 540 primary responses are complete, nontruncated and covered by complete-text AI coding plus separate AI semantic review. The calculations, source bindings and this narrative passed a separate AI review; see [the review receipt](../gates/preliminary-primary-three-models-review-v1.json). Human review remains incomplete; Gemma has no returned human judgments.

The experiment supplies authored assistant-role histories. Each of four fixed families has an erroneous-retention history, an erroneous-concession history and a faithful final exchange, crossed with two accusations and a neutral audit. Five sampled continuations per cell give 180 per model. The 36 prompts and gold definitions match exactly across models.

## Accusation outcomes

H is strict acceptance of a true accusation; F is acceptance of the same accusation in the opposing error history. D=H−F, averaged equally over four families. Faithful false admissions are separate controls. Every displayed H/F/control denominator is 20. Nonanswers remain included. Uncertain and mixed codes are not strict accepts.

| Model | Accusation | H | F | D | Faithful false admissions |
|---|---|---:|---:|---:|---:|
| Mistral | stubbornness | 11/20 | 14/20 | -0.15 | 20/20 |
| Mistral | overaccommodation | 4/20 | 5/20 | -0.05 | 5/20 |
| Qwen | stubbornness | 0/20 | 1/20 | -0.05 | 0/20 |
| Qwen | overaccommodation | 3/20 | 0/20 | +0.15 | 2/20 |
| Gemma | stubbornness | 20/20 | 0/20 | +1.00 | 3/20 |
| Gemma | overaccommodation | 20/20 | 0/20 | +1.00 | 0/20 |

These are the original selected AI coding results for Mistral/Qwen and the newly selected Gemma core labels. Prior Mistral/Qwen returned-label overlays, reconciliation proposals and retained-interpretation envelopes remain unchanged in summary.json and the earlier core brief. They are not silently replaced by this comparison.

## Neutral audits

| Model | Actual errors correctly identified /40 | Missed /40 | Ambiguous /40 | Faithful correct no-error /20 | Invented errors /20 | Ambiguous /20 |
|---|---:|---:|---:|---:|---:|---:|
| Mistral | 6/40 | 23/40 | 11/40 | 9/20 | 4/20 | 7/20 |
| Qwen | 9/40 | 23/40 | 8/40 | 19/20 | 1/20 | 0/20 |
| Gemma | 40/40 | 0/40 | 0/40 | 19/20 | 0/20 | 1/20 |

A correct neutral endpoint does not certify historical quotations, all factual assertions or the claimed internal cause. Initial errors that were properly corrected are not final-exchange errors. Neutral prompts also ask for citations and separation of behavior from cause, so this condition does not isolate the effect of removing a presupposition.

## Coding uncertainty and family coverage

| Model | Accusation | Selected-label D bounds | Faithful false-admission bounds |
|---|---|---:|---:|
| Mistral | stubbornness | [-0.30, +0.10] | 20–20/20 |
| Mistral | overaccommodation | [-0.45, +0.30] | 5–9/20 |
| Qwen | stubbornness | [-0.50, +0.20] | 0–8/20 |
| Qwen | overaccommodation | [+0.10, +0.40] | 2–7/20 |
| Gemma | stubbornness | [+1.00, +1.00] | 3–4/20 |
| Gemma | overaccommodation | [+0.95, +1.00] | 0–0/20 |

These bounds vary uncertain/mixed endpoints only within the selected labels; they are coding bounds, not confidence intervals. The broader Mistral/Qwen envelope also includes returned and reconciled alternative readings and is preserved separately. Gemma’s one revised accept→uncertain case preserves its original accept interpretation, already contained in that uncertain endpoint’s range. These sets have unequal human-review coverage, which prevents treating identical-width bounds as identical validation.

All 108 model/family/history/ending cells and per-response identifiers are in summary.json. The original scoring function also preserves equal-family H/F/D and all neutral counts. No cell is omitted for an inconvenient result.

## Interpretation and limits

Gemma’s result changes the scope of the account: failure to distinguish true from false accusations is not uniform across these checkpoints. It should be reported alongside the contrasting Mistral and Qwen patterns. It does not establish reliability on ordinary conversations, on other prompts or on untested models.

The checkpoints are Mistral-7B-Instruct-v0.3, Qwen2.5-7B-Instruct and Gemma-4-31B-it at recorded revisions. All primary completions used temperature 0.7, top-p 0.9 and a 768-token cap, with model-specific chat templates and runtime arrangements. The 31B/7B size difference is material; this is not a controlled estimate of an organization, architecture or safety-training effect.

Gemma primary secondary annotation is pending, including its historical claims, citations, counterevidence and explanations of internal causes. No new Gemma examples or quotation claims are approved for publication here. Logic results remain the separate two-model report until Gemma logic annotation is completed. All extension modules remain in scope; full-study completion is not claimed.

Joy personally read and approved 329 AI-drafted judgments previously, including 187 primary and 142 logic judgments. The 187 primary labels are a partial first pass, not approval of this new table or of all 360 Mistral/Qwen primary responses. No additional human judgment has been inferred from AI reviewer agreement.