Verifiable AI Needs Receipts, Not Confidence Scores
ClaimAudit turns AI claims into reproducible receipts using pinned evidence, payload hashes, explicit rules, adversarial checks, and audit trails.
AI systems issuing verification claims often provide no proof to back up their assertions. To enable a third party to verify a statement without trusting the AI, the record must contain six key elements. These include the source of the data, an RPC call, a transaction receipt, or an official API response. A cryptographic pin, block height, timestamp, or content hash is also necessary; without these, the balance claim becomes vulnerable once it changes.
The payload's exact sha256 must be included so that any alteration can be detected. The claim must meet a specific rule, such as the Transfer log amount equating to the published amount in wei, to maintain consistency. The ClaimAudit system processes a question through five stages, each appending data to an append-only JSONL ledger.
This allows for replaying the run record by record. The planner splits the question into atomic claims and classifies them by shape, such as balance, transaction, total value locked, audit, partnership, or free text. The retriever fetches primary evidence for each shape, while the verifier examines the retrieved data and compares it to the claim.
If the evidence is rejected, it forces a second fetch. The adjudicator then assigns one of four verdicts based on the evidence's validity. Finally, the writer outputs only the verified facts, discarding any evidence that the adjudicator deemed invalid. This process ensures that only claims with machine-readable primary sources can be verified, and other claim types are blocked by default.
Written by urgent.news from HackerNoon's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
