EvidenceRegistry v0.3.0 — retaining AI-agent review evidence without treating the agent’s verdict as authority
I've released EvidenceRegistry v0.3.0, a Rust implementation for evidence-sensitive AI-agent review workflows.
The problem I wanted to separate is fairly simple:
an agent saying CLEAN, a process exiting 0, retained output bytes, structural validity, and a policy-derived decision are not the same thing.
EvidenceRegistry retains exact Records and Journal state, binds a review Result to a specific Request and target, and exposes a bounded Store-owned Review Admission path without treating structural validity or the agent's own explanation as semantic truth.
v0.3.0 also includes the separate AI Agent Evidence Binder companion. Its public workflow can create a Request before dispatch, retain managed attempts and bounded outputs, produce a Result for the original Request, derive the applicable local Admission where the retained inputs permit it, and inspect the retained attempts again after reopening the workspace.
Hermes is the first real-agent integration/dogfood target, but the Rust Core and read-only Journal CLI do not depend on Hermes.
The exact Binder implementation tree was natively qualified on Windows x86_64, Linux x86_64, and macOS arm64.
There are deliberate limits: this is not a correctness oracle, execution authentication system, general authority engine, sandbox, deployment authorization layer, or proof that an AI reviewer was honest or found every defect.
The boundary I keep coming back to is:
STRUCTURALLY_VALID != AUTHORITATIVELY_VALID
and even a Policy acceptance is not semantic truth.
v0.3.0 is currently a source-only release under MIT OR Apache-2.0.
Repository:
I'd mainly be interested in this being useful as a reference implementation or design point for people working on agent review, provenance, evidence retention, and reproducible decision workflows.