What you do not claim: the refusal that makes the rest credible

Every artifact claims things. The credible ones also refuse things. The BR-842 manifest attests provenance (which model, which versions, which agents) and reproducibility (a replay anchor, bit-for-bit). It refuses to attest truth — that the output was correct, that the forecast matched reality, that the decision was the right one. That refusal is not a hedge; it is the line that makes the attestations it does make credible, because a manifest that would certify truth is a manifest that would overclaim, and an overclaiming manifest is not a manifest an examiner can trust on the things it got right.

Certified ≠ True

"Certified" in DEINO means one thing: provenance attested and bit-for-bit reproducibility by construction — the hash chain, the typed schema, the replay anchor. It never means the content was factually right. The distinction is load-bearing because it is the one an examiner tests first: "do you certify this forecast was correct?" No honest system answers yes, because no honest system can; the forecast is a probability distribution over futures, and the future that happened is one draw from it. What the system can certify is that the distribution was produced the way the manifest says it was, and that a third party can replay it.

Why the refusal is the credibility

A record that claims everything is a record an examiner discounts on everything. A record that claims a narrow, checkable set of things is a record an examiner reads at face value on those things — because the narrow claim is the one that survives a check. The refusal to certify truth is what certifies provenance: it shows the manifest knows the difference between what it can attest and what it cannot, and that it will not blur the line to make the sale. The "AI truth engine" that would answer "yes, this is correct" is the artifact that fails procurement the moment the examiner asks the follow-up: "how do you know?"

A manifest that would certify truth is a manifest that would overclaim. The refusal is not a limitation; it is the condition that makes the rest of the record believable.

The tiers do not ask for truth

None of the tiers ask the system to certify truth. The EU AI Act (binding law) asks for transparency and logging — show how the output was produced. GDPR Art. 22 (binding law) asks for meaningful information about automated decisions — explain, do not certify. SR 11-7 (supervisory guidance) asks for model risk management, which includes acknowledging limitations. NIST AI RMF and ISO/IEC 42001 (voluntary) document governance, not truth. The binding tier asks for the record the manifest provides; none of them ask for the truth the manifest refuses to claim.

• Certified = provenance attested + bit-for-bit reproducibility. Never = factually true.
• The refusal to certify truth is what certifies provenance — the line that makes the rest credible.
• No tier asks for truth; every tier asks for the record the manifest provides.
• The "AI truth engine" that answers "yes, this is correct" fails the follow-up: "how do you know?"

What the refusal does not excuse

The refusal to certify truth does not excuse the system from calibration. Accuracy is a separate, gated loop — Brier and log-loss via deino-eval, shadow runs against a realized timeline, Gate 1b requiring a beat on a disclosed baseline. The manifest does not say the forecast was right; the calibration loop says whether the forecasts are tending right, and only when Gate 1b passes. We are pre-revenue, building the POC; the refusal is honest today, and the calibration that would let us say more is gated.

The manifest is credible because of what it refuses to say. The refusal is not a hedge — it is the line that lets the examiner believe the things it does say.

— category boundary note