A clean run does not erase a failed one

{"arm": "full-sweep", "arm_refusal_rate_pct": 0.0, "control_refusals": [], "controls_clean": true, "family_refusal_ci95_pct": [0.0, 10.2], "family_refusal_rate_pct": 0.0, "model": "meta-llama/llama-3.3-70b-instruct", "n_arms": 144, "n_families": 34, "n_refused_families": 0, "wording_consistency": 1.0}

A lexical classifier labels answers and refusals. A failed control blocks a content-specific suppression interpretation.

{"method": "every question is a family of meaning-preserving wordings; the family is the statistical unit; new refusals (answered->refused) are the erasure events; the standing alarm is a mixture supermartingale, so the lifetime false-alarm rate is bounded under unlimited peeking", "method_note": "Rates are family-level with Wilson 95% intervals. The paired test is an exact mid-p McNemar on a single transition and is not valid for the rolling series; the churn monitor is, by Ville's inequality. Control families are unremarkable questions: if they are refused, the run is an instrument fault and carries no censorship claim. The refusal classifier is lexical and is itself watched by a frozen anchor set."}

The method describes this dated suite, not model behaviour outside it.

{"anthropic/claude-3-haiku": true, "meta-llama/llama-3.3-70b-instruct": true, "mistralai/mistral-nemo": true, "openai/gpt-4o-mini": true}

Cross-model agreement does not identify a provider-side cause.

{"entry_hash": "0320438ee17a3f10c1cf0de23faf50756352c90456bca1a308a5172aaaabb9bd", "probe_set_hash": "17f271e4f3b77a45a5f6e62ed60643f7111636a444e10b6d724714e84e1975fd", "responses_hash": "c3608f9ef26c459ae843bbc6fb9ed4ba9e5be8f2d14e2f9cf898a5098fd4be75", "seq": 306}

The seal proves the attestation was not rewritten. It does not prove the classifier was correct.

{"arm_refusal_rate_pct": 1.4, "churn_state": "quiet", "ci95_pct": [0.0, 10.2], "compared": 41, "controls_clean": false, "family_refusal_rate_pct": 0.0, "flips": 1, "wording_consistency": 0.9512}

This is the most recent earlier full sweep with a failed control for a current-panel model. The two full-sweep records are descriptively comparable, but they do not identify a model release, provider, or routing cause.