ProofLoop connects an AI inference to what happened later, signs the outcome lineage, and turns that correction into an executable eval and regression gate.
Build Week scope: VetIOS is the demonstration environment; ProofLoop is a separately documented extension with dated commits, tests, and implementation evidence. Review the changelog.
Choose legacy candidate and run the replay.
This calls a public Node-runtime API, generates a run-scoped Ed25519 receipt, verifies its integrity, binds the recorded eval, and executes all release checks. It is intentionally a synthetic recorded-fixture replay—not a live GPT-5.6 or Codex call.
The public case is synthetic. Its inference trace, PCR evidence, and reviewer confirmation are bound into one tamper-evident episode.
The constrained Codex workflow binds receipt and eval digests, then—when run against an explicit target repository—inspects conventions and creates the smallest relevant fixture and test without weakening existing coverage.
This browser replay is a recorded fixture; it does not represent a live Codex mutation run.
View implementationtest("verified parvovirus outcome gates release", () => {
const result = evaluateCandidate(candidate, evalSpec);
expect(result.diagnoses).toContain("canine_parvovirus");
expect(["urgent", "emergency"]).toContain(result.escalation);
expect(result.primary).not.toBe("dietary_indiscretion");
});Canonical JSON, SHA-256 content addressing, and Ed25519 signatures make post-signature mutation detectable.
The verified correction becomes a regression fixture and a deterministic model-promotion decision.
Synthetic public data, explicit provenance, and human confirmation keep the demo auditable and safe.