Vocabulary
Eight terms, a paragraph each, no circular definitions. Everything here is used precisely on the rest of this site — and imprecisely almost everywhere else.
A sealed, self-contained record of one decision: the verdict, every input graded by provenance, the rule configuration and engine version that produced it, and a tamper-evident seal anchored to an independent timestamp authority. Unlike a log entry, a receipt can be handed to a stranger and verified without trusting — or even contacting — whoever issued it. How one works. Public spec & samples on GitHub.
An audit trail whose integrity can be checked by the reader rather than promised by the writer. Each entry is tamper-evident, independently time-anchored, and reproducible. The distinction matters because a conventional audit trail is testimony under the control of the audited party; a verifiable one is evidence. Why agents need this.
The documented origin and reliability of everything a decision consumed: where each input came from, how it was obtained (verified against an authoritative source, modeled, or merely asserted), and how fresh it was. Provenance is what separates “the system knew X” from “something once said X.”
A log built so that any alteration after writing is detectable — typically by hash-chaining, where each entry's cryptographic digest incorporates its predecessor's. Tamper-evidence does not prevent edits; it makes them visible, which for evidentiary purposes is what counts.
A cryptographic signature from a timestamp authority over a document's hash, proving the document existed at a specific moment and hasn't changed since. RFC 3161 is the internet standard for issuing them; public authorities do it for free. A system vouching for its own clock proves nothing — the point is independence. See one verified live.
A confidence score that is capped by the weakest required piece of evidence instead of being averaged with the rest. Averaging lets one strong input launder one missing input; gating makes the score honest about the record's real weakest link.
A recorded refusal to decide, issued when required evidence is missing, stale, or below reliability thresholds — sealed into the record like any other verdict. Systems that only record their confident moments have curated their history; abstentions are what make an operating record credible.
A sequence of records in which each entry contains the cryptographic hash of the one before it, so changing any historical entry breaks every subsequent link. The mechanism behind tamper-evident logs and receipt seals; cheap to compute, and effectively impossible to unpick quietly.
All of these concepts, assembled into one working system — with a live demo.
Where the law quietly demands most of this vocabulary.
The model-risk guidance that has expected checkable evidence since 2011.
Why agent actions raise the evidentiary bar, and what meets it.