Aegis v2.5.3

Highlights

• Reconciled all ten documented agentic benchmark scenario classes with deterministic workflow-quality fixtures and the three checked-in controlled replay samples.
• Separated deterministic coverage, controlled replay references, live eligibility, and actual pass evidence so fixture presence can no longer be reported as evaluation success.
• Added a four-tier evaluation contract for deterministic static checks, controlled transcript replay, opt-in repeated held-out live evaluation, and sampled blind human review.
• Kept the current implementation deliberately bounded to deterministic checks and single-transcript development replays. Repeated live runs, variance, held-out execution, and blind review remain contract-only follow-up work.

Comparative Evaluation

• Defined exact current arms for baseline-no-aegis and aegis-auto.
• Kept previous-aegis conditional on future candidate revision evaluation rather than making it a universal runtime arm.
• Added a versioned advisory JSON report with top-level and per-sample overallPass, structured failures, arm outcomes, score deltas, preserved unknowns, and a non-authoritative promotion status.
• Prevented a passing comparison delta from hiding an arm contract failure.

Validation And Safety

• Added independent matrix and replay validators for mappings, tiers, arm structure, expected pass semantics, promotion limits, and unsupported evidence claims.
• Added adversarial negative coverage for coordinated fixture drift, duplicate or missing arms, premature previous-aegis use, live or held-out overclaims, and automatic promotion.
• Restricted generated replay workspaces and reports to strict children of the repository-local .tmp directory, including a sentinel regression proving the .tmp root cannot be deleted.
• Hard-cut the benchmark matrix and replay manifest to schema version 2 without retaining a legacy compatibility path.

Verification

• bash tests/e2e/run-all.sh --full --host-profile fast
• Aggregate suites: 4/4
• Layer 1: 40/40
• Layer 2: 6/6
• Layer 3: 6/6
• Version audit: all nine declared package and host manifest surfaces are synchronized at 2.5.3, with no undeclared version references.

Release Boundary

This release remains Aegis Method Pack (runtime-ready). It preserves multi-host plugin-installable distribution and does not add authoritative GateDecision, PolicySnapshot, candidate promotion, or completion authority.