Evidence about AI agent actions that anyone can check again from the bytes.
Replay a public conformance corpus through the bundled reference verifier, with nothing to clone. Python 3.11+ and uv are enough:
uvx agent-evidence-vectors==0.18.0Add --verifier "<your command>" to test your own verifier against the same records.
Check any verifier of signed AI-agent and supply-chain records against a public corpus. Choose a tool by task.
Explore retained runs or contribute one in the Open Evidence Lab. Browse the task catalog in HTML or Markdown. Read the agent guide for commands and pinned sources. Use the replay procedure to run a published example and keep its evidence.
| Task | Project |
|---|---|
| Choose a tool or inspect recorded claims and retained runs | Agent Evidence Conformance Board |
| Sign and verify DSSE envelopes in Rust | dsse |
| Reject duplicate JSON keys and canonicalize in Rust | jcs-admit |
| Record file actions or check native delegation evidence | Probity Observer |
| Test a verifier against conformance cases | Agent Evidence Vectors |
| Install a pinned verifier and check an evidence claim | Probity Verify |
| Use shared definitions for agent evidence claims | Probity Vocabulary |
| Run a checkout composition and replay its retained claims | Checkout composition reference |
| Check artifact digests or evaluate admission policies | Probity Admission |
| Check an agent guard against replay and regression cases | Probity Ratchet |
Ratchet turns each observed agent failure into an executable guard.
It proves the guard by replay against recorded history and fresh held-out runs before admitting it, and signs
every admission decision with the
agent change admission predicate.
Install it with pip install probity-ratchet==0.1.0; its
README
gives the commands and the recorded results.
Find release states, source revisions and pinned install commands in the Atlas project catalog.
Try a worked task:
Each project explains its supported commands and examples in its README. Verify links to source-pinned installation and checked task recipes. Admission includes policy-engine rules and a pre-call budget API and CLI for bounded local workloads. Retained runs keep source revisions, readers, operators and rerun instructions with the result.