Isolated Agentic Benchmark Program
- Added a version 4 agentic benchmark contract with 30 cases across ten
scenario classes, pairedbaseline-no-aegisandaegis-autoarms, and
explicit development, standard held-out, and extended held-out profiles. - Added deterministic scheduling with bounded workers, paid-attempt ceilings,
infrastructure retry limits, persistent wall-clock accounting, crash-safe
resume behavior, and preservation of every paid attempt result. - Added Bubblewrap-based execution isolation, provider preflight, sealed auth
handling, bounded process supervision, confidential artifact cleanup, and
fail-closed exposure checks.
Observable Scoring And Reproducible Reports
- Added observable-outcome scoring that is independent of arm vocabulary,
validates exact workspace effects, and keeps scorer unknowns separate from
infrastructure failures and contract failures. - Added controlled replay, immutable verification inputs, matrix and case
validators, sanitized Markdown/SVG report projection, and adversarial checks
against report, path, metric, and publication-boundary drift. - Kept benchmark results advisory: development evidence cannot be promoted into
held-out or general product claims, and report output does not grant runtime
authority or completion authority.
Deeper-Cause Proof And Progressive Disclosure
- Strengthened systematic debugging so root-cause closure requires falsifiers,
causal-topology evidence, recurrence analysis, quick-exit proof, and
topology-specific anti-disguise checks when deeper-cause signals are present. - Progressively disclosed debugging and verification detail into directly
routed references while preserving the executable hot path, escalation
triggers, compact completion receipt, and current defaultTDD Mode: off. - Kept the quick bug lane proportional while requiring an explicit
change-necessity and owner-fit decision before source edits.
Reliability And Review Hardening
- Added shared secure atomic JSON writes, output and artifact limits, symlink
boundary checks, structured destructive-tool classification, exact changed
path contracts, and stronger semantic benchmark assertions. - Closed independent review findings around confidential evidence boundaries,
extended-profile wall budgets, interrupted wave persistence, and Linux CI
sandbox provisioning. - Kept the repository root free of branch-local
docs/aegis/work records and
excluded local.codex/configuration from the release surface.
Verification And Release Boundary
- The formal release gate remains
bash tests/e2e/run-all.sh --full --host-profile fast, with the Linux-only
benchmark supervisor covered by the repository CI environment. - Bumped all declared package and host manifest versions directly from
2.5.3
to2.5.5, following the selected release version, with no undeclared
version references found by the version audit. - This release remains
Aegis Method Pack (runtime-ready), preserves
multi-host plugin-installable distribution, and does not add authoritative
GateDecision,PolicySnapshot, benchmark promotion, or completion
authority.