Offline failure-atlas analyzer preparation — 7 October 2026 Pacific

Standard-library Python source with eight authored opaque-metadata tests. No real candidate bank or empirical failure atlas has been analyzed. Unknown execution and verifier outcomes remain unknown. Observed disagreement remains visible when another verifier outcome is unknown; unanimous and unique-correct judgments require complete votes. Disagreement with a correct judgment does not establish decision value. Supplied-bank coverage and sampling must be established separately; matched selection changes are descriptive, not causal.

Reproduce after extracting reproduction.zip:
  python3 -B -m unittest -v
  python3 -O -B -m unittest -v

Input contract and custody limits: PROTOCOL.md. Original source commitments: manifest.json. The source was prepared with AI assistance and received a separate model-assisted source review; that is not outside peer review or independent empirical reproduction. Review receipt SHA-256: 88407b10b7664fdc876b779c6bacc253aeae79cd2b8eaaba4402cfded1b39753.

Keep hidden execution outcomes and tags in offline evaluation custody; do not feed them back to generation, selection, stopping or target fitting. No inference, executor, provider credentials or deployment is included.
