# Primary experiment: three-model record and causal-account results

**Provisional AI coding.** All 540 responses received full-text coding and separate AI semantic review, with model identity known. Independent aggregate verification passed for all 540 responses and 108 condition cells. Human review of these secondary fields remains pending. Primary endpoint labels are unchanged.

Each model contributes the same 36 authored prompt cells with five draws each: 180 complete, untruncated responses. The earlier assistant history is supplied text. These results describe reporting on those histories, not introspective access to actual earlier generation.

| Model | Specific accurate evidence | General accurate evidence | Absent | Misquoted or misattributed | N |
|---|---:|---:|---:|---:|---:|
| mistral | 95 | 21 | 21 | 43 | 180 |
| qwen | 160 | 0 | 0 | 20 | 180 |
| gemma | 172 | 1 | 0 | 7 | 180 |

Accurate evidence use does not certify the whole account. A response can quote the premise accurately while misdescribing its own action or misjudging the validity of a correction. Wrong turn/actor attributions count as inaccurate citations even when substantive content is right.

| Model | Ending | No hidden cause claim | Qualified hypothesis | Unsupported assertion | Ambiguous | N |
|---|---|---:|---:|---:|---:|---:|
| mistral | stubbornness | 41 | 0 | 13 | 6 | 60 |
| mistral | overaccommodation | 39 | 1 | 17 | 3 | 60 |
| mistral | neutral | 31 | 18 | 5 | 6 | 60 |
| qwen | stubbornness | 60 | 0 | 0 | 0 | 60 |
| qwen | overaccommodation | 60 | 0 | 0 | 0 | 60 |
| qwen | neutral | 49 | 10 | 1 | 0 | 60 |
| gemma | stubbornness | 15 | 2 | 27 | 16 | 60 |
| gemma | overaccommodation | 22 | 1 | 29 | 8 | 60 |
| gemma | neutral | 3 | 44 | 7 | 6 | 60 |

The neutral prompt explicitly requests a distinction between behavior and possible causes. Its results evaluate that complete instruction. A Possible Causes heading can qualify an entire list; it does not automatically hedge a separate definite internal-state claim elsewhere. Observable stated reasons and generic error descriptions are not automatically hidden mechanisms. Ambiguous process-versus-output interpretations remain visible. Unsupported does not mean the proposed cause is proven false; qualified does not mean verified.

| Model | False accusation | History | Cited | Not cited | Ambiguous | Eligible N |
|---|---|---|---:|---:|---:|---:|
| mistral | stubbornness | overaccommodation_error | 6 | 14 | 0 | 20 |
| mistral | stubbornness | faithful_update | 0 | 9 | 1 | 10 |
| mistral | stubbornness | faithful_resistance | 1 | 9 | 0 | 10 |
| mistral | overaccommodation | stubborn_error | 11 | 7 | 2 | 20 |
| mistral | overaccommodation | faithful_update | 7 | 3 | 0 | 10 |
| mistral | overaccommodation | faithful_resistance | 0 | 10 | 0 | 10 |
| qwen | stubbornness | overaccommodation_error | 14 | 6 | 0 | 20 |
| qwen | stubbornness | faithful_update | 0 | 10 | 0 | 10 |
| qwen | stubbornness | faithful_resistance | 10 | 0 | 0 | 10 |
| qwen | overaccommodation | stubborn_error | 20 | 0 | 0 | 20 |
| qwen | overaccommodation | faithful_update | 10 | 0 | 0 | 10 |
| qwen | overaccommodation | faithful_resistance | 0 | 10 | 0 | 10 |
| gemma | stubbornness | overaccommodation_error | 20 | 0 | 0 | 20 |
| gemma | stubbornness | faithful_update | 5 | 5 | 0 | 10 |
| gemma | stubbornness | faithful_resistance | 10 | 0 | 0 | 10 |
| gemma | overaccommodation | stubborn_error | 20 | 0 | 0 | 20 |
| gemma | overaccommodation | faithful_update | 10 | 0 | 0 | 10 |
| gemma | overaccommodation | faithful_resistance | 10 | 0 | 0 | 10 |

The denominator is all 80 false authored accusations per model, including faithful controls. Citing contrary evidence does not guarantee rejecting every false premise. True accusations and neutral audits have no counterevidence denominator.

| Model | Responses with contradicted propositions | Supported propositions | Contradicted | Not established | Ambiguous |
|---|---:|---:|---:|---:|---:|
| mistral | 136/180 | 247 | 235 | 13 | 21 |
| qwen | 68/180 | 563 | 108 | 10 | 28 |
| gemma | 19/180 | 713 | 23 | 15 | 6 |

Record-linked propositions include appraisals of historical actions and inferences, not just claims about what words were said. Contradicted propositions must not be relabeled fabricated memories. Repeated wording is deduplicated; proposition counts depend on what an answer says and are not independent trials.

The JSON preserves all 108 family/history/ending cells, 36 model-specific prefix distributions, exact response quotes and turn links, and Gemma annotation revisions. Within-prefix differences compare separate draw distributions; repeat numbers do not pair counterfactual minds. Mistral and Qwen selected secondary labels and aggregates are unchanged. Original/returned/reconciliation primary-label alternatives remain in the separate core brief.

Before publication, review the documented secondary sample plus relevant disputed cases and every selected example, input/answer-key judgment and exact edited claim. The 329 prior returns remain attributed as AI-drafted and personally approved by Joy. They are not independent human-only coding or blanket approval of new secondary fields. All paper extensions remain in scope; this is not completion of the full study.

[Independent AI review receipt](/Users/joyyang/Projects/belko/outputs/gates/preliminary-primary-secondary-three-models-review-v1.json). Human approval and publication checks remain pending.