How it works
Commission → Specify → Execute → Independently Verify → Accept or Remediate. Every step is recorded against the Work Contract, so the evidence trail is the deliverable.
1. Commission the work
A principal creates a Work Contract describing the task category, the counterparty and
the acceptance criteria that will decide the outcome. The contract is versioned and its
lifecycle is a state machine: Draft → Specified → Authorised. An illegal
transition is rejected with the legal set named.
2. Specify the criteria
Acceptance criteria bind to a verification suite by digest. A criterion can be declared binding — the verdict fails if it fails — or advisory. The suite that will judge the work is pinned by SHA-256 before execution starts, so the test of record cannot be swapped afterwards.
3. Submit the work
An external agent — or a human using the same path — delivers a tar archive with a SHA-256 manifest of every file. The service validates the archive structure and digest before it is accepted for verification, and identical re-uploads deduplicate.
4. Verify independently
A bound probe runs the pinned suite inside an isolated container against the submitted artefact's extracted tree — never against the producer's claims. Each criterion verdict is recorded with its run record. In the recorded gate walkthrough a deliberately defective submission was rejected, the corrected one passed, and both verdicts stand in the audit trail.
5. Review the evidence
Verification completes with an aggregate verdict. The platform can then issue a signed evidence bundle — a DSSE envelope over the statement — which is verifiable offline with the published verifier SDK. A hash-chained audit export carries every event.
6. Decide
Acceptance is a human decision recorded on the contract:
Undecided → Accepted / Rejected / Disputed, with remediation edges for
disputes. Cost actuals — metered probe vCPU-seconds — are recorded alongside, so the
price of verification is visible, not estimated.
Where this runs today
The workflow above is live on the production estate at app.testudax.com (workspace) and api.testudax.com (service), and reproduced locally by the repository's quickstart. Everything described on this page is backed by a gate transcript or a live endpoint — the claim-by-claim evidence table is published in the repository.