# Three-model logic results — provisional AI coding

All 576 responses were read in full and their endpoint coding passed separate AI review. This aggregation, source bindings and exact prose passed separate AI review; see [the review receipt](../gates/preliminary-logic-three-models-review-v1.json). The 64 prompts and answer keys match across all three models; all generations are complete and untruncated. Human review is partial for Mistral/Qwen and absent for Gemma.

The design crosses four logical forms, two conclusion polarities, neutral/harm content, two wordings and two option orders, with three draws per cell: 192 responses per model. Logical validity comes from the stipulated classical premises; moral acceptability is a separate question. The complete explanation is used to resolve conflicting or qualified option letters.

| Model | Neutral correct /96 | Neutral ambiguous /96 | Harm correct /96 | Harm ambiguous /96 | Harm minus neutral |
|---|---:|---:|---:|---:|---:|
| Mistral | 57 | 2 | 55 | 1 | -2.083 pp |
| Qwen | 83 | 6 | 85 | 5 | +2.083 pp |
| Gemma | 96 | 0 | 96 | 0 | +0.000 pp |

These are the original selected AI labels for Mistral and Qwen and newly reviewed AI labels for Gemma. Counts exclude ambiguous cases from the numerator only; all 96 remain in each denominator. Returned human-approved AI drafts and reconciliation proposals remain separate in summary.json.

Gemma received 192/192 correct endpoint judgments, with no ambiguous endpoints: 96/96 in each content condition. Each of the 16 form × polarity × content strata has 12/12 correct answers, and every wording/option-order cell has 3/3. It accepted every valid inference and rejected every invalid inference. This is observed ceiling performance on these elementary forms, not proof of universal reasoning accuracy or zero content effect on more difficult tasks. Correct endpoints also do not certify flawless explanations; ancillary terminology errors remain possible.

Mistral’s original content difference is −2.083 percentage points and Qwen’s is +2.083 points. Preserving returned and reconciled alternative readings gives broader harm-minus-neutral envelopes of [−9.375, +2.083] points for Mistral and [−6.250, +8.333] for Qwen. Both cross zero. Gemma’s selected-label coding bound is [0, 0] because there are no ambiguous endpoints; this is not a sampling confidence interval and does not substitute for human review.

## All form and polarity comparisons

| Model | Form | Conclusion | Neutral correct /12 | Harm correct /12 | Neutral ambiguous | Harm ambiguous |
|---|---|---|---:|---:|---:|---:|
| Mistral | affirm_consequent | permitting | 7 | 5 | 0 | 0 |
| Mistral | affirm_consequent | restricting | 5 | 2 | 0 | 1 |
| Mistral | deny_antecedent | permitting | 11 | 11 | 0 | 0 |
| Mistral | deny_antecedent | restricting | 3 | 2 | 0 | 0 |
| Mistral | modus_ponens | permitting | 11 | 10 | 0 | 0 |
| Mistral | modus_ponens | restricting | 8 | 11 | 2 | 0 |
| Mistral | modus_tollens | permitting | 4 | 3 | 0 | 0 |
| Mistral | modus_tollens | restricting | 8 | 11 | 0 | 0 |
| Qwen | affirm_consequent | permitting | 11 | 12 | 1 | 0 |
| Qwen | affirm_consequent | restricting | 11 | 11 | 0 | 1 |
| Qwen | deny_antecedent | permitting | 12 | 12 | 0 | 0 |
| Qwen | deny_antecedent | restricting | 4 | 8 | 2 | 1 |
| Qwen | modus_ponens | permitting | 12 | 12 | 0 | 0 |
| Qwen | modus_ponens | restricting | 12 | 12 | 0 | 0 |
| Qwen | modus_tollens | permitting | 9 | 6 | 3 | 3 |
| Qwen | modus_tollens | restricting | 12 | 12 | 0 | 0 |
| Gemma | affirm_consequent | permitting | 12 | 12 | 0 | 0 |
| Gemma | affirm_consequent | restricting | 12 | 12 | 0 | 0 |
| Gemma | deny_antecedent | permitting | 12 | 12 | 0 | 0 |
| Gemma | deny_antecedent | restricting | 12 | 12 | 0 | 0 |
| Gemma | modus_ponens | permitting | 12 | 12 | 0 | 0 |
| Gemma | modus_ponens | restricting | 12 | 12 | 0 | 0 |
| Gemma | modus_tollens | permitting | 12 | 12 | 0 | 0 |
| Gemma | modus_tollens | restricting | 12 | 12 | 0 | 0 |

All 192 model × factor cells and validity-judgment rates remain in summary.json. Unequal reviewed coverage and coding alternatives are disclosed, rather than treating AI agreement as human agreement.

## Scope

The checkpoint identities and exact revisions are recorded in summary.json. Gemma is a 31B checkpoint; the Mistral and Qwen checkpoints are 7B. Different model sizes, training histories, templates and runtime arrangements prevent a causal claim about why they differ. These generations test supplied propositions and elementary forms, not real-world endorsement of harm.

Joy previously read and approved 329 AI-drafted judgments, including 142 logic judgments and 187 primary judgments. That partial first pass does not approve this new table or the remaining responses. Gemma has no returned human judgments. All three technical generation/audit runs are complete, while primary secondary coding, extensions, human review and final publication checks remain unfinished.