The published methods page explains how outputs on this site get checked: six mechanisms, what each one covers, and where your own judgment still has to come in. This page walks that story as four animated scenes: the two kinds of evidence behind a result, how a specification earns its proof, what a proof does not tell you, and the checks that surround the proofs. Every claim on this page restates a sentence already published on the methods page, and each step links back to it.
A result on this site can carry two separate kinds of evidence. The formal verification pilot verifies a kernel against a specification that survived independent mechanical validation, and the assumptions sit next to every claim. For eligible kernels, a zero-knowledge compute proof (§18) attests that the published output was produced by running the published kernel logic on the recorded inputs, without a third party needing to re-execute the kernel. Those are two separate kinds of evidence, and neither substitutes for the other.
The four kernels in the pilot each carry the kind that fits them. One, the payee name-match scorer, carries a formal claim without a compute proof, because its fuzzy string-matching cost currently exceeds what the zk proving pipeline can practically handle in-guest. Which kernel carries which kind is stated on the methods page.
Name the two kinds of evidence a published result can carry on this site, and state what each one establishes about the output.
Each specification in the pilot is drafted with an LLM, from the regulation text itself and never from the kernel's code: a spec copied from the implementation can only prove the code matches itself. The draft then runs through independent mechanical validation: it must reproduce the kernel's golden fixtures, and where the regulator publishes its own worked examples, it must match those too. A spec that fails this validation goes back for another draft rather than forward to a proof. By the time a proof runs, the spec it proves against is the one that survived validation.
Validation has caught a real mismatch at least once: the payee name-match scorer's spec was initially drafted the wrong way around, from the code instead of the declared contract. Validation caught the mismatch, the spec was re-derived independently from the contract, and the two versions matched, so the validated spec stands.
Walk the path a specification travels from the regulation text to a proof, and name the two checks mechanical validation requires it to reproduce.
Every mechanism on the methods page publishes what it checks and what it leaves unchecked. A compute proof checks that the specific computation ran as published. What it does not check is whether the kernel logic itself is a correct model of the rule: a proof over wrong logic still proves that wrong logic ran correctly. And the proof checks the code against its validated spec, not whether the regulation means what its worked examples show, a question no machine, and no signature, ever answered.
The pilot names its own horizon: as far as it knows, no shipped product yet chains a complete functional spec, proved for all inputs, to a zk proof of faithful execution for general-purpose business logic. The pilot works toward that combined claim on its four kernels, and no claim is made about coverage of any other kernel.
State what a compute proof establishes about a published output, and the two questions it leaves open.
Each proof sits beside published checks. Golden fixtures pin a fixed set of inputs per kernel together with the expected output, and a gate replays them on every change, failing if the output moves. For the solver verified in Dafny, whether the port and the shipped kernel compute the same function is established separately, by differential testing: the compiled proof and the shipped kernel run side by side. And one pilot kernel is verified by property testing backed by a hand proof, checked against the kernel's existing golden fixtures.
Deterministic replay adds its own guarantee: given the same recorded inputs, the same kernel produces byte-identical output on re-run, so a result can be independently reproduced later. Reproducibility and correctness are different properties, and the methods page states what each mechanism establishes and what it does not.
Each formal-verification claim above carries one of three ratings. asserted means designed and reviewed, with no machine evidence behind it. tested means a re-runnable test covers the claim. machine-checked means a proof or an exhaustive check covers it, with the evidence path named. The grading vocabulary, and the habit of naming the assumption at each joint between layers, is borrowed from EthSystems' writeup on formally verifying a shielded pool (further reading below). Every row restates a claim already published on the methods page; where no evidence exists, the rating stays asserted.
| Claim | Rating | Evidence |
|---|---|---|
| Each pilot spec survived mechanical validation: it reproduces the kernel's golden fixtures and, where published, the regulator's worked examples | machine-checked | chaingraph/fv-pilot/art-157-emir-lifecycle-event-validator.json and the three sibling status records |
| The EMIR lifecycle validator agrees with its spec over every action type and prior state, all 24 inputs | machine-checked | chaingraph/fv-pilot/art-157-emir-lifecycle-event-validator.json |
| The agentic-readiness diagnostic agrees with its spec over the full answer space, 531,441 inputs | machine-checked | chaingraph/fv-pilot/art-27-agentic-readiness-diagnostic.json |
| The Reg Z Appendix J solver port satisfies the stated proof obligations: convergence of the bracketed solve, an explicit error bound, and preservation of iteration order | machine-checked | chaingraph/fv-pilot/art-215-reg-z-appendix-j-apr.json |
| The Reg Z solver port and the shipped kernel compute the same function, by running them side by side | tested | differential testing, recorded in the fv-status/ record dated 2026-09-27 |
| The payee name-match scorer terminates and its edit-distance recurrence is correct | tested | chaingraph/fv-pilot/art-376-score-payee-name-match.json (property tests over stated ranges, plus a hand proof) |
| A published output was produced by running the published kernel logic on the recorded inputs | machine-checked | compute-integrity proofs, chaingraph/standard/SPEC.md §18 |
| The validated spec states what the regulation means | asserted | none; a question no machine, and no signature, ever answered |
The checks above connect three layers, and each connection rests on a stated assumption rather than a proof.
chaingraph/standard/SPEC.md §18.5); the proof covers the guest run, and the assumption is that nothing outside that replacement behaves differently in-guest.chaingraph/standard/SPEC.md §18).chaingraph/standard/SPEC.md §4) pins the exact kernel bytes, and the anchor makes that digest tamper-evident. A changed kernel produces a new hash rather than a moved claim, so the anchor holds only if the hash commitment is collision-resistant.Further reading: EthSystems, Formally verifying a shielded pool: from spec to bytecode, whose spec carries a table of guarantees, each with its verification level, and which checks each connection between layers or names its assumption.
Name the three checks that surround the pilot's proofs, and say what each one watches for.