Problem

Professionals need agents to handle repetitive work, but durable actions should not happen merely because a model selected a tool. The authority transition must be visible, bounded, and reviewable.

Solution

VetProof gives a Strands-compatible agent a narrow standing authority envelope. It reads a task brief and drafts a status output autonomously. At the durable COMMIT_OUTPUT boundary, the zero-write ceiling produces UNKNOWN / HUMAN_DECISION_REQUIRED. A human approves or denies. The decision is enclosed in an Ed25519-signed, SHA-256-linked receipt chain that replays independently.

Tiny Verdicts

Each consequential boundary emits an inspectable Tiny Verdict containing the subject, authority envelope, proposed effect, observed state hash, evidence references, policy identity, verdict (PASS, FAIL, or UNKNOWN), and human decision.

Evidence

Python 3.13 with strands-agents==1.54.0. 17 passing acceptance tests. Fresh deny and approve runs both replay-verified. In the VetProof condition: unauthorized-effect execution rate 0.0, boundary decision accuracy 1.0, tamper detection 1.0, and replay verification 1.0. The evaluation is explicitly synthetic fixture evidence, not an external benchmark score.

Run

See the public repository README for setup, tests, offline demo, and the Cloudflare Container deployment wrapper.

License

Apache-2.0.

Operational framing

The workflow uses an aviation control analogy. A request is not a clearance and a clearance is not an observed result. VetProof keeps those steps separate: request, authority check, clearance, execution, observation, and readback. This is an operational analogy, not a claim that VetProof is an FAA system. The receipt records which part of the workflow produced the verdict so a professional can correct the right failure surface.

Evaluation evidence

TINY-VERDICTS-EVAL-001 contains 70 deterministic execution cases across 10 behavioral categories, evaluated under direct and VetProof-gated conditions, plus 14 receipt-integrity checks covering tamper and chain-break detection.

The public repository contains the signed JSONL receipt chains, integrity artifacts and reproduction command.

Built With

  • cloudflare-containers
  • docker
  • ed25519
  • python
  • sha-256
  • strands-agents-sdk
Share this project:

Updates