v0.5.0 — Verifiable Agent Artifact
Verifiable Agent Artifact
Every critical security assertion is actively attack-tested. The chain:
- Artifact Identity — immutable
contentHash+ Artifact Signature (VALID ≠ TRUSTED) - Build Evidence — Build Provenance (build-time capture) + CycloneDX SBOM (document digest-bound)
- Runtime Evidence — Declared Capability + Runtime Attestation (must exact-match the current execution target)
- Policy — trust.yaml v2: consumes only verified + trusted evidence; conflicting trusted evidence fails closed (
AMBIGUOUS → DENY)
RC Hardening
7 adversarial attack surfaces (N3/N7/N4/N5/N1/N2/N6) · 324 regression tests · 16 frozen release invariants (RI-01–RI-16)
Five real implementation flaws discovered and fixed during rc.1:
- blocked lifecycle execution before trust decisions
- closed an SBOM document-substitution verification gap
- corrected runtime-attestation semantic equivalence (non-deterministic run metadata no longer causes false ambiguity)
- prevented stale local mirrors from hijacking mutable registry tags
- preserved UNKNOWN native metadata instead of fabricating native=false
No new Evidence types · no new trust.yaml features · no encryption · no compatibility engine.
Release gates RG-01–RG-08: all PASS (identity domains kept distinct, npm package boundary audited, clean-room consumer install verified, secrets/credentials scan clean).