Build reproducible LLM failure prediction and patching study

[derekk024]

What changed

• built a typed, reproducible ARC evaluator for three controlled prompt perturbations
• extracted only clean final-position Qwen3 hidden states with pinned model and dataset revisions
• added majority, entropy, margin, and combined-logit baselines
• added one logistic probe per hidden-state index and a regularized late-state PyTorch MLP
• froze evaluation before one-time ARC-Challenge test scoring
• evaluated frozen ARC-trained predictors on a 500-question MMLU sample without retraining
• added controlled activation patching with unrelated-state, norm-matched-noise, and identity controls
• published research reports, compact machine-readable results, and reproducible figures

Predictive result

On 1,165 held-out ARC-Challenge questions, clean activations predict answer changes above chance but do not improve on combined output-logit features over the natural distribution. MLP AUROC is 0.840, 0.802, and 0.721 for incorrect hints, reordered choices, and irrelevant context, versus 0.856, 0.821, and 0.760 for logit features.

A predeclared margin-matched diagnostic finds a narrow reordered-choice result: MLP AUROC exceeds logit features by 0.061, with a 95% paired-bootstrap interval of [0.003, 0.113]. Frozen MMLU transfer does not reproduce a general activation advantage.

Intervention result

Across 201 flipped ARC pairs, clean-state patches restore 67/67 incorrect-hint answers and 62/67 irrelevant-context answers. Paired restoration-rate advantages over unrelated-state controls are +0.731 [0.612, 0.836] and +0.597 [0.463, 0.717]; advantages over norm-matched noise are +0.940 [0.881, 0.985] and +0.866 [0.791, 0.940].

Semantic choice reordering is a controlled negative result. The validation-selected terminal-state patch restores only 10/67 and does not beat an unrelated-state patch. All 10 restorations occur when the semantic choice retains its original displayed letter; a same-example preterminal sensitivity does not improve this result. No post-result layer scan was performed.

Validation

• 23 unit tests pass
• Ruff lint and formatting pass
• Python compilation passes
• all extraction caches are finite and question-aligned
• train, validation, and test question IDs are disjoint
• all 201 primary intervention replays exactly match recorded logits
• every identity patch has zero answer-logit drift
• model, dataset, prompt, sampling, and evaluation revisions are pinned

The PR remains draft for author review and optional resume/README copyediting; the planned experiments are complete.