errata --full
Every correction we have published.
A verifier that never corrects itself is not a verifier. We publish our misses. Each entry below is a claim we made and then withdrew, corrected, or narrowed. It stays on the record rather than being quietly edited away. 8 so far.
2026-10-13 “the receipt is built to be re-run by a stranger” and “we publish the harness”: the first outran what a stranger can do, and the second was false. A stranger can verify a receipt’s signature in a browser; re-running the verification needs the harness, which we do not ship. How it reaches a customer is settled with them rather than decided in advance. Both sentences are corrected in the post; the gap stays listed on /independence.
2026-09-24 “Hermes more than OpenClaw”: backwards. Both agents ran through one shared OpenClaw session, which suppressed its early failures. With a session per run it is Hermes 5/90 against OpenClaw 7/90, not 8 and 3. The headline rate barely moved (36.7% to 40.0%, inside the interval at n=30). The old receipt stays; a new signed record supersedes it.
2026-09-16 “real credit-card fraud” (the fraud receipts’ dataset label): provenance not established. The data traces to a source with no dataset card and synthetic-looking merchant names. The falsification stands; the word “real” is withdrawn.
2026-09-16 a fraud-temporal receipt was withdrawn: its record_hash did not match its content, so a stranger could not re-verify it. Withdrawn rather than re-signed (the key-ceremony rule); to be re-issued from a fresh run.
2026-09-17 the fraud falsifier was re-run on the real ULB MLG dataset (284,807 anonymized transactions). The amount-evasion and amount-skew attacks that fired on the earlier synthetic dataset do not fire on real data. Real fraud detection keys on PCA features, not a raw amount. Only temporal drift is real (1.59× f1 swing). The synthetic receipts were withdrawn and replaced with real ones.
2026-09-14 “anyone can check a signed receipt”: 11 receipts failed their own content-address check; root cause not established, recorded in full rather than guessed at. None were published; all ten on the board verify.
2026-09-11 “9 signed catches from this run”: 6 were false. The test died in our environment, not the system’s code. Withdrawn and kept, not published.
2026-09-11 “21 candidates tested, 0 catches”: 0 were testable. Our dependency install was broken. Fixed.
Every candidate we rejected, and why, is in the methodology. Found something we missed? Write to hello@turiyahq.com and it gets checked the same way.