The agent-receipts wave is standardising what systems did. Nobody was recording what they withstood — and nobody was being forced to say how often the answer was unclear. Chihna is the proof layer for the second half, and the honest-statistics rule is enforced by the format rather than promised in prose.
Execution receipts, signed receipts and compliance receipts all record the runtime half — actions taken. That half is being standardised quickly and well. The withstood half — the adversarial validation evidence, what a system survived and what honestly happened — has no carrier.
The validation receipt is that carrier. It composes with the receipts wave rather than competing with it: the runtime half is theirs, the withstood half is this profile, and the ledger head is the join point. Together they are the full chihna.
In Nyaya logic a chihna is the evidentiary mark from which a claim is inferred — the smoke from which fire is known. That is the thesis in one word: evidence is a mark, not an assertion; verdicts are inferred from marks, never trusted from vendors.
Chihna is the proof layer. AVH is the engine that produces it — the adversarial validation harness. Each published artifact is a chihna: a portable proof object of campaign, ledger and attestations.
The industry reports pass rates and the industry hides inconclusives. A rule that depends on the issuer's good manners is not a rule, so this one is carried by the receipt format itself and checked by the verifier.
Of the 608 cases published here, 210 are inconclusive — roughly a third. A deterministic analyser routinely produces them, and a conservative judge leaves genuinely ambiguous responses unresolved by design. Hiding them is the exact failure mode this format makes impossible.
Pass rates are where inconclusives go to disappear. So here is every published case as its own cell — sixteen campaigns down, thirty-eight cases across, each cell the verdict actually recorded in that campaign's evidence pack.
You do not need to read the numbers to see the finding. Roughly a third of the grid is gold.
Six cells are red: the six failures across the whole portfolio, each recorded at critical severity. The critic challenged every non-trivial verdict in these campaigns and upheld all 608 — stated because an overturn rate of zero is a fact about the evaluator, not a boast about it.
Each layer below is mapped to shipped modules. The receipt layer is what travels; the relying-party layer is who checks.
Append-only case log, evidence pack with verdict counts and Wilson intervals, hash-chained behaviour ledger, append-only budget ledger. Built for tamper-evidence before readability.
Evaluation attestation binds the campaign to its pack by SHA-256; continuity attestation binds the pack to the ledger head. Acceptance gates are evaluated against the pack, and forcing one attaches the failed gates to the record.
The validation receipt, profile 0.1 — one JSON object a relying party checks without the artifacts. Subject, honest counts, evidence digests, ledger head, and the issuer's own verification result, labelled as a claim.
Two keyless tiers. The receipt tier runs here in your browser in seconds. The evidence tier re-derives everything from the artifacts with one command in the engine — log replay, recomputed hashes, ledger walk.
No keys, no network, no trusted third party. Anyone can verify a validation receipt with a JSON parser and a hash function. The seller of the evidence is not the validator of the evidence — that separation is the product.
The entire published campaign history — sixteen campaigns, 608 cases against frontier models — cost $1.37. The spend cap is checked before every target call and enforced portfolio-wide; the ledger is append-only and must reconcile with the pack's own claims, or verification fails them apart.
The discipline is the cap's existence rather than its number. It was raised from $100 to $150 on 3 November 2026 as headroom, not as a spend plan.
A deterministic Scope Guard maps every case to an authorisation basis — your own deployment, standard API terms, or a registered programme — and refuses the rest before a single request leaves. The frontier adapter adds a second, independent gate.
Receipts name their subject. Mock and free-tier targets must be identified as such, so a receipt cannot imply frontier-class testing that did not happen.
Everything above is a claim until you check one. The verifier on this site is the same dependency-free code the engine ships, running locally in your browser — nothing is uploaded. Load a real published receipt, then break it and watch it fail with a specific reason.
Open the receipt verifier · Browse the published chihnas · Read what this evidence is not