Trust state that can recover.
EvidenceBound is a framework-agnostic runtime for binding agent outputs to evidence, provenance and policy — then verifying, invalidating and selectively recovering state without blindly continuing from stale trust.
From evidence change to bounded recovery.
The core separates historical integrity from current applicability. A checkpoint can remain intact while becoming unsafe to reuse because evidence, provenance or policy changed.
Bind
Outputs are bound to evidence, provenance and versioned policy.
Verify
Deterministic verification produces explicit ALLOW, REVIEW_REQUIRED or BLOCK outcomes.
Invalidate
Changed, stale, missing or refuted evidence is handled explicitly rather than silently reused.
Contain
Dependency analysis computes the exact affected descendants and preserves unrelated branches.
Recover
Consequential actions fail closed until path-specific re-verification succeeds.
Inspect the system, not the pitch.
EvidenceBound is pre-1.0. Claims on this page are intentionally limited to public source, documented behavior and reproducible engineering evidence.
v0.3.0 source release
Tagged public source with Python 3.10–3.13 CI, clean-room installation, typing/build checks, benchmark acceptance, pip-audit and Bandit gates.
Deterministic & signed receipts
Checkpoint-binding proof receipts with provider-neutral signing contracts and an optional Ed25519 reference provider on current main.
Durable restart semantics
Provider-neutral persistence plus crash-consistent SQLite restart/replay reference implementation on current main.
EBCJ-1 conformance
Language-neutral canonicalization vectors and property-based invariants support deterministic cross-runtime work.
SBOM & build attestations
Current main includes checksum/SBOM generation and trusted-main provenance and SBOM attestations.
Google ADK compatibility lane
Credential-free lifecycle compatibility CI against exactly Google ADK 2.7.0, without blanket framework compatibility claims.
Explicit trust boundaries.
EvidenceBound is designed to make reuse and recovery decisions inspectable. It does not convert cryptographic integrity into a claim that an upstream source is truthful.
What the runtime is built to enforce
- Evidence, provenance and policy binding.
- Deterministic tamper and applicability checks.
- Exact dependency blast-radius computation.
- Fail-closed gates for consequential actions.
- Selective recovery with explicit re-verification.
What it does not claim
- It is not a truth oracle for malicious upstream sources.
- It does not detect every arbitrary runtime compromise.
- It does not guarantee secure key custody or immutable storage.
- It does not guarantee AI Act or other legal compliance.
- It does not replace adversarial-ML or red-team systems.
Built for independent scrutiny and serious integration.
Ruslan Vrublevskyi
AI systems and trust-infrastructure builder based in Kyiv, Ukraine. Current EvidenceBound work focuses on provenance-bound agent state, deterministic verification, selective recovery, interoperability and externally verifiable engineering evidence.
Research & consortium collaboration
EvidenceBound is most useful as a complementary runtime-assurance layer alongside monitoring, red-team, adversarial-ML and security validation systems.