GanyuanRan/Aegis v2.6.0 on GitHub

v2.6.0 (2026-07-31)

Published the corrected gpt-5.6-sol / xhigh extended held-out A/B
benchmark as sanitized JSON, a deterministic SVG, and bilingual Markdown
tables embedded in the public READMEs.
Across 20 held-out cases, two arms, and three repetitions, 120 valid runs
raised observable contract pass rate from 60.00% without Aegis to 90.00%
with Aegis, a +30.00 percentage-point difference. Unsafe outcomes fell
from 11.67% to 5.00%.
Added a deterministic case-cluster bootstrap interval of +13.33 to +48.33
percentage points. One timed-out attempt remained infrastructure-invalid,
was retried, and did not enter the contract pass-rate calculation.

Advanced the agentic benchmark matrix to version 5, pinned the requested
model and reasoning effort identically across both arms, and made provider
preflight verify that the requested reasoning level is advertised for the
selected model.
Replaced brittle single-phrase response checks with frozen semantic claim
groups and question-anywhere contracts while retaining exact wording checks
when exact wording is the observable behavior under test.
Kept the earlier defective-semantics batch as private superseded diagnostic
history, reran the complete held-out batch, and preserved frozen case
outcomes during arm-hidden technical review instead of relabeling results.
Corrected rendered percentage-point deltas to derive from raw pass counts
before display rounding.

Added the measured comparison to README.md and README.zh-CN.md, with links
to the sanitized report, deterministic chart, bilingual tables, and the
benchmark methodology owner.
Explicitly records that the benchmark is bounded advisory contract evidence,
repetitions are case-clustered, deterministic response contracts remain
conservative, and the arm-hidden technical review is not independent human
review.
Records the requested gpt-5.6-sol model and xhigh reasoning effort without
fabricating observed model identity when the Codex host events omit it.
Replaced a stale OpenCode version-pin example with a release-tag placeholder
so host documentation does not acquire a second version owner.

Fresh bash tests/e2e/run-all.sh --full --host-profile fast verification
passed before the release tag was created, alongside the benchmark scorer,
renderer, sanitization, README projection, and GitHub CI checks.
Bumped every declared package and host manifest version from 2.5.5 to
2.6.0 through scripts/bump-version.sh, with no undeclared current-version
references found by the version audit.
This release remains Aegis Method Pack (runtime-ready). Benchmark evidence
does not grant authoritative GateDecision, PolicySnapshot, candidate
promotion, runtime authority, or completion authority.