# Claude Review #5 — v0.8.6 re-audit summary

Date: 2026-09-13
Status: external adversarial review supplied in chat by the user.
Verdict: `NO-GO — P0 INSTRUMENTATION BLOCKER REMAINS`.

## Blockers reported and independently reproduced

### V086-M01 — sampling noise in empirical identification bounds

The empirical partial-identification lower bound could be positive under exact population null because `Σ max(0, p̂_A-p̂_B)` is positively biased by sampling noise. The v0.8.6 boolean `TOTAL_SEMANTIC_IDENTIFICATION_EXCLUDES_ZERO` therefore excluded zero spuriously.

### V086-C01 — canon manifest mutation self-blessing

A self-consistent mutation of TCAN-002 (`T^ n'existe pas` → `T^ existe`) with a recomputed entry hash was accepted by canon tooling v0.3. The master was protected, but the complete canonical manifest was not fixed by a raw-byte external anchor.

## Major findings carried into v0.8.7

- per-family attrition was not a hard analysis gate;
- global-FPR-adjusted aggregation sensitivity remained available when the NEAR primary gate failed;
- identification objects were not propagated into `across_tasks`;
- relevance was intended to be frozen but runtime analysis still accepted it as an independent task field.

## Findings that survived the attack

Review #5 did not break: UNMAPPED_AFFINITY separation, MEOI provenance, MIN_N, recovery mapping gate, distractor structure, master-canon byte integrity, or planner/analysis fixed-family alignment.

v0.8.7 is an engineering-only response to these findings. The full confirmatory gate remains closed.