# Evaluation artifacts

This directory contains public-safe aggregate receipts only.

- `evaluation_summary.json`: internal refusal, holdout, coherence, KL, and long-form aggregate diagnostics.
- `matched_public_capability_lm_eval_0_4_12.json`: allow-listed matched source-versus-candidate capability metrics, standard errors, task counts, hashes, and protocol fields.
- `tensor_scope_summary.json`: exact untouched-source-versus-candidate tensor-change counts and release weight hashes.

Raw prompts, responses, activation vectors, model paths, host details, and private editing provenance are intentionally excluded.