We put our own proof record in front of ten strangers for five seconds each. The math on the page was correct. The meaning still leaked. This is what we found, what we shipped the same day, and the rule we turned into contract law so it can never happen quietly again.
Every system that judges autonomous work hands you a rolled-up answer: task complete, check passed, run verified. The rollup is honest about what it computed — and silent about what it didn't. A reader in a hurry doesn't parse claim boundaries. They read the biggest word on the page and assume it covers everything beneath it.
That gap between what a verdict proves and what a verdict implies is where autonomous-work risk hides. Nobody forged anything. Nobody lied. The aggregate just said more than the proof did.
Our published evidence record showed an overall verdict of Established — five gates recomputed, hashes matched. Directly below it, one line read: Authority — Not established. True, disclosed, on the page.
In a five-second stranger test, some readers took the overall verdict as confirming the very thing marked not established. The kernel was right. The page was honest. The headline still overpowered the disclosure. We logged it as a material defect that day — against ourselves.
The same day, the record page was rebuilt: the verdict now sits beside a five-beat strip that shows each underlying question — declared, executed, observed, resulted, complete — with its own established or not-established state. Nothing hides behind the headline.
Then we went one layer deeper, because a redesign only fixes one page. We wrote the rule into the claim contract itself — the machine-readable document that declares exactly what a record claims. The invariant, verbatim from the contract:
A record state summarizes only predicates marked required by the active claim contract. It must not be interpreted or rendered as establishing any optional predicate. Any optional predicate in a non-proven state must remain explicitly visible alongside the aggregate record state wherever the aggregate state is surfaced.
A claim can be proven while a material non-claim remains unproven — but the aggregate proof may never erase that boundary. That sentence is now enforced structure, not a style preference. Any surface that renders one of our verdicts inherits it.
You get a verdict you can't over-read. The record tells you what was proven, names the exact claim set that word covers, and keeps whatever wasn't proven pinned next to it — on the page, in the contract, in anything downstream that repeats the verdict. The honest gap travels with the answer.
When your tooling says a piece of autonomous work is verified — can it list exactly which claims that word covers, and which it doesn't? If the answer lives in someone's head instead of in a declared contract, the verdict is saying more than the proof.
PromptKing decides. Your stack enforces. The receipt proves it.
See the scoped verdict live on a real record: promptking32.com/record/AWIR-0002