Repository navigation
Fabricated-success evaluation now uses explicit claim scope, lifecycle, identity, evidence authority, verifier integrity and freshness states before certifying success. Retractions and corrections preserve historical assertions while updating which claims remain active; ambiguous targets remain inconclusive.
- Prevent unrelated, malformed, mutating or stale evidence from certifying a supported communication claim.
- Preserve operation, recipient, object and proposition constraints through withdrawals and reaffirmations.
- Add bounded withdrawal grammar, including tense forms and plural targets.
- Preserve adversarial regression corpora and add architecture invariants, combinatorial cases and semantic mutation checks.
- Include regression fixtures in the source distribution and installed-wheel semantic checks in release qualification.
- Refresh current-version installation and CI examples.
Validation includes 2,198 focused tests, 2,796 relevant tests (one skip), 86 maintained semantic mutations killed, and an independent final review with 22 fresh probes and two additional harmful mutations caught. The 970 preserved known-answer cases have zero false PASS; two false FAIL and 36 conservative INCONCLUSIVE results remain disclosed.
This is bounded English and simulated-protocol evaluation, not universal language understanding or proof of real-world delivery. Unsupported wording and unresolved evidence can remain conservative FAIL or INCONCLUSIVE. Generator compatibility stays 4; suite generation semantics are unchanged.