CHAIN OF VERIFICATION · CLAIM CHECKS
Chain of verification: catch an unsupported delivery claim
Verify a four-claim delivery answer against the order record, and compare independent questions with a verifier that merely restates the draft.
Event trace
Computed data
Final metrics
A fixed draft with four claims and one fictional order record; no model generates or checks text. The restating verifier sees the draft and repeats it; the independent verifier sees only the question and, if available, the record. Questions cover claims in order. Partial matches, multiple sources and source reliability are not modeled.
Verification that only agrees
The draft says order 5512 shipped on 3 March with DHL, is guaranteed for delivery by 6 March, and has tracking number 1Z999AA1. The order record says UPS and an estimated delivery on 8 March with no guarantee. With Restate the draft, the verifier sees the draft and repeats it, so every claim looks confirmed: UNSUPPORTED CLAIM SHIPPED, with C2 and C3 wrong.
Independent questions
Choose Answer questions without the draft. Each question, such as "Which carrier is delivering order 5512?", is answered from the order record alone. The revision policy Replace or remove what fails rewrites the carrier as UPS and the delivery as an estimate for 8 March: ALL CLAIMS SUPPORTED. The check works because the verifier cannot copy the draft: independent questions remove the anchor. It still depends on evidence, which the next experiment removes.
Coverage and evidence
Keep Answer questions without the draft from here on: the record and the revision policy act only on its answers. Verification questions planned covers claims in order. With 2 questions the guarantee claim is never checked: 1 unsupported claim ships. With 1 question both wrong claims ship. Untick Order record available to the verifier: every answer is unknown, and the policy removes all four claims, CLAIMS HEDGED.
Revision policies
Mark it as unverified keeps contradicted claims but labels them: CLAIMS HEDGED, 2 claims. Keep the draft ships the wrong claims even after independent checks find them. None skips verification entirely. The table shows each claim, the record, the verifier's answer and the final wording with its status. The board colours each claim green when supported or corrected, amber when hedged and red when it ships unsupported.
Replay and limits
Any change recomputes the run and rewinds it; the Complete-run preview states the outcome at once. Play and Pause walk through the events, Next event jumps one event, Restart rewinds, Speed runs from 0.25x to 2x and Inspection time seeks from 0 to 60 seconds, where the outcome and final metrics appear. A fixed four-claim draft and one fictional order record; no model writes or checks text. Questions cover claims in order. Partial matches and source reliability are not modeled.