HONESTAS stops a decision when the models disagree, when the signer lacks the authority, when the evidence moves, when it cannot tell whose data it is, when time runs out and when a record has changed. Each stop has a code, a plain reason and a next step, and each one is kept as a record.
A check that never says no proves nothing. These are the six that stop an insurance decision in HONESTAS, in the words the product uses, with what happens next.
GATE_DISAGREEMENTA second, independent model from a different family worked the decision out again and reached a different result.
What happens nextNothing goes out. Both answers go back to the room side by side, the chair decides, and the disagreement stays in the record.
AUTHORITY_EXCEEDEDThe amount or the kind of decision is outside what this person's delegated authority allows, on the version of the authority in force today.
What happens nextHONESTAS names who can sign it and opens a referral to them. Nobody can raise their own limit.
EVIDENCE_MOVED_BEFORE_COMMITA paper arrived or changed between the signature and the seal, so the signature no longer covers everything in the record.
What happens nextThe seal waits and the signer is asked to look again. Their second signature covers the full record.
SCOPE_REFUSEDA request did not say which firm it belongs to, whose data it is or which unit is asking.
What happens nextNothing is read. The request comes back with the missing line named, and the refusal is recorded.
DEADLINE_EXCEEDEDA model, a room or an approval did not finish inside the time set for it.
What happens nextThe case moves to the next person on the escalation line, with everything so far. A wait that runs out never approves itself.
HASH_MISMATCHA record's fingerprint no longer matches its seal, or a re-run of a validation does not reproduce the sealed result.
What happens nextThe copy is refused as evidence and the verifier names the record. The original, still sealed in the record store, is the one that counts.
Every code appears in the product, in the record and in the API's answers. See developers for how your systems receive them.
A refusal is not an error message that scrolls away. It is sealed like any decision, with the case, the reason, both sides of the disagreement and who was told, so an auditor can count them and read each one.
Refusal · CLM-2026-5120SampleThese are not options in a menu. No customer, administrator or Adverant engineer can turn them on for one firm.
How a caller sounds can bring a person onto the call sooner. It never reaches a price, an underwriting decision, a claim or a fraud check, and staff are never analysed for emotion. The AI Act already bans AI that infers emotions at work, except for medical or safety reasons, under Article 5(1)(f).
No claim is declined, no cover refused and no policy cancelled by a model on its own. A named person with the authority signs every decision that goes against a customer, and their reasons go in the letter.
Nobody can change a sealed record, raise their own limit or switch a check off without it showing. An override is a new decision, with a name, a reason and its own seal, linked to the one it replaces.
A refusal appears on the screen where the work is, in plain words, with the step that resolves it and the person who can take it.
The seal, the scope check and the second-model check cannot be switched off. Authority limits and deadlines are yours to set, and changing them needs a second person.
A person with the authority can make the decision. That is a new decision with their name and reasons, linked to the refusal, which stays in the record.
A refusal sends the case to the right person with everything gathered so far, so nobody starts again. Most decisions never meet one; the ones that do are the ones a supervisor would ask about.
No. Customers see plain words in their letter or app. Codes are for your teams, your auditors and your systems.
Bring a claim you think AI should not settle alone. We will run it and show you the refusal, the record and the person it goes to.