Highlights
- Publish the matrix-v7 standard snapshot for the frozen Aegis 2.10.8 build: 44 valid outcomes over 22 cases, with one observation per arm and case. The observed contract-pass rate was 59.09% → 86.36%, and unsafe outcomes were 18.18% → 4.55%. These are advisory 2.10.8 results, not a 2.10.9 performance or cross-version improvement claim. The documented non-blind review exception was accepted; the maintainer did not independently review raw outputs. (#66)
- Reject benchmark retry opt-ins from standard and extended main-track evidence; improve process-cleanup CI stability, OpenCode managed skill refresh, and Pi/OMP fast-path routing. Strengthen correction of inherited false completion claims and the confirmation boundary for distributed-code retirement. (#72)
- Add future-only benchmark checks for actual confirmation, active-handoff correction, and concrete pre-edit rationale. The published matrix-v7 contracts and scores remain frozen. (#73)
- Synchronize eight package and plugin manifests at
2.10.9and link the newer benchmark snapshot from the English and Chinese READMEs. The older extended snapshot remains labeled with its own run profile.
Verification and remaining work
- The release branch passed
bash tests/e2e/run-all.sh --full --host-profile fast: four layers passed, zero failed. The version audit found all eight manifests in sync and no undeclared2.10.9strings. - #67 and #68 remain open for repeated current-build behavioral evidence. Previously failing cases will be replayed diagnostically after this exact tag; a single replay cannot establish that intermittent failures have disappeared. Independently prepared held-out cases are needed for a stable improvement claim.
Aegis remains Aegis Method Pack (runtime-ready) without runtime or completion authority.