Highlights
- Publish the first matrix-v7
standard-held-outbenchmark snapshot for Aegis 2.10.1: 44 valid outcomes across 22 cases, one observation per case and arm. It is advisory evidence, not a 2.10.8 performance or cross-version improvement claim. The flagged results received contributor review with arm labels visible; the maintainer matched the ledger to frozen scores without independently inspecting raw outputs. (#57) - Fix the benchmark renderer's false credential alert on known case IDs and make the event scorer accept
scopedchange rationale in future batches. Previously published scores were not recalculated. (#57, #62) - Update the DeepSeek Harness bootstrap message to use the producer-owned
plugin:aegissource kind required by session format v4 in DSH 0.1.7-rc.1. DeepSeek Harness remains a developer-preview host without a maintainer-verified release-level live smoke verdict. (#63) - The original reporter confirmed that the v2.10.7 Python 3.10.12 updater fix resolved the issue. This release adds no further updater change.
Verification
- All eight package and plugin manifests are synchronized at
2.10.8; the version audit andgit diff --checkpass. - Exact-head Ubuntu CI passed for
c6e4eba, including the full host-neutral aggregate, Python 3.10 updater suite, OpenCode base suite, and Codex plugin sync. - Local DeepSeek Harness deterministic checks and scoped updater tests passed. The isolated DSH install check was skipped because
dshis not installed locally. - The four committed benchmark bundles re-rendered byte-identically on Linux. Windows Git Bash passed 41/42 Layer 1 checks, including representative Codex host smokes; its one local failure is the Linux-only offline benchmark fixture.
Aegis remains Aegis Method Pack (runtime-ready) without runtime or completion authority.