EvidenceBound Core v0.3.0 is the first independently installable alpha release of the framework-agnostic EvidenceBound trust, provenance, deterministic verification, and selective-recovery runtime for autonomous AI systems.
Included in this release:
- deterministic EBCJ-1 canonicalization and evidence/provenance/policy binding;
- historical-integrity/current-applicability separation;
- tamper-evident proof receipts with explicit security limits;
- fail-closed ALLOW / REVIEW_REQUIRED / BLOCK outcomes;
- exact dependency-graph blast-radius invalidation and selective recovery;
- replay/idempotency guards and provider-neutral structured events;
- plain Python integration and a dependency-free Google ADK adapter seam;
- Python 3.10-3.13 CI, clean-room install, strict typing/linting, build/wheel checks, pip-audit and Bandit security gates.
Acceptance evidence for the tagged commit includes 32 tests, golden acceptance, and the reproducible synthetic selective-recovery scenario: 100 full-restart nodes versus 25 recompute / 75 reusable / 75 recomputations avoided. These benchmark counts are scenario-specific, not a universal performance claim.
EvidenceBound does not claim that a digest makes an upstream source truthful, does not treat changed evidence as automatically refuted, and does not trust persisted historical state merely because it was previously stored. Signed receipts, crash-consistent persistence, maintained real-framework compatibility CI, external security review, and independent external adoption remain explicit follow-on work.
License: Apache-2.0. Unconditional runtime dependencies: none.