Human Control Plane for Verifiable AI Agents
EvidenceBound is early-stage open-source AI safety infrastructure for binding agent actions to evidence, provenance and executable policy, with deterministic verification, bounded blast radius and recoverable human control.
The core design principle is simple: an AI system should not be able to grant itself trust. EvidenceBound keeps verification, policy enforcement and human authority outside model-generated claims and makes unavailable, stale or mismatched evidence visible rather than silently promoting it to certainty.
| Repository | Purpose |
|---|---|
| evidencebound-core | Framework-agnostic runtime for evidence/provenance binding, deterministic verification, policy enforcement and fail-closed selective recovery. |
| evidencebound-recovery-mesh | Trust-aware selective recovery for autonomous agent fleets after a trust break. |
| evidencebound-authority-cut | Reversible autonomy with human authority kept outside the model-callable surface and correction propagated through affected descendants. |
| evidencebound-releaseproof-dws | Differential reverification for changed document evidence under a frozen historical review-equivalence policy. |
- Evidence visibility - missing, stale, blocked or mismatched evidence remains explicit.
- Provenance - claims are bound to inspectable inputs and transformations.
- Executable policy - trust decisions are enforced outside natural-language model output.
- Deterministic verification - machine-checkable gates decide whether evidence and state remain acceptable.
- Bounded blast radius - trust breaks invalidate only work that depends on the broken evidence or authority.
- Recoverable human control - human grants, revocations and corrections remain outside model self-authorization.
- Website: evidencebound.org
- Maintainer: Ruslan Vrublevskyi
- Contact:
ruslan@evidencebound.org - LinkedIn: Ruslan Vrublevskyi
EvidenceBound does not claim that cryptographically authenticated provenance makes source data true, that an LLM can certify its own behavior, or that a working prototype is equivalent to a security-reviewed production system.