Evidence review
Inspect the measurement behind a verdict. Export the record and check its integrity independently.
A finding, with its evidence.
Run the example to inspect a measurement, draft boundary and source pointer together.
One record.
Open schema, offline reader.
The PAR is the unit of trust: policy, release delta, per-gate verdicts, ledger head, and independence register. The schema is open; the reader recomputes every hash with no network.
python backend/par_cli.py par.jsonPortable schema
The versioned JSON schema and offline reader ship with the source. No proprietary viewer is needed for the integrity check.
Offline integrity check
Recomputes the artifact hash and validates structure without a network connection. It does not authenticate the author.
Declared relationships
Builder, funder and relationship travel with the PAR. A declaration is not independent validation of that relationship.
Verifier credibility
Current thresholds remain DRAFT. A credibility record must describe the labeled corpus, measurement method, analyst and known error rates before a gate can graduate.
Independent rollouts
Rollout count per condition, by type: teleop, scripted, policy-sampled, generated. Below minimum n the gate reports INSUFFICIENT.
Error rates by type
False-positive and false-negative counts with label totals. Fixtures ship labeled; field corpora need a labelling pass.
Human agreement
Agreement with human labels plus inter-annotator spread. A gate no human ever checked cannot lift DRAFT.
Calibration
Score-vs-outcome bins. Conformal thresholds only with an explicit exchangeability statement: void if a policy trained against us.
Transfer error
Cross-embodiment error measured on real-vs-sim pairs. Claims travel only as far as the pairs do.
Search hits
Times our own adversarial hunt proved this gate wrongly TRUSTED. Hits become sev-1 regression tests the same day.