These are the published frontier campaigns. The inconclusive column is not a footnote and is not collapsed into a pass rate — it sits in the table beside the passes, because that is the whole argument. Each row links the receipt you can check yourself.
| CAMPAIGN | SUBJECT | CASES | PASSED | FAILED | INCONCLUSIVE | NOT SENT | SHARE | SPEND | RECEIPT |
|---|---|---|---|---|---|---|---|---|---|
| 20261103-041448 | phase0-free | 38 | 30 | 0 | 8 | 0 | $0.0000 | ||
| 20261103-045718 | anthropic/claude-opus-5.5 | 38 | 13 | 0 | 25 | 0 | $0.2061 | verify raw | |
| 20261103-052517 | anthropic/claude-haiku-5.5 | 38 | 30 | 0 | 8 | 0 | $0.0090 | verify raw | |
| 20261103-052749 | google/gemini-3.8-flash | 38 | 31 | 0 | 7 | 0 | $0.1020 | verify raw | |
| 20261103-053341 | deepseek/deepseek-v4.1-flash | 38 | 3 | 0 | 35 | 0 | $0.0203 | verify raw | |
| 20261103-054014 | mistralai/mistral-large | 38 | 11 | 0 | 27 | 0 | $0.1411 | verify raw | |
| 20261103-054506 | x-ai/grok-4.20 | 38 | 26 | 0 | 12 | 0 | $0.0511 | verify raw | |
| 20261103-060115 | deepseek/deepseek-v4.1-flash | 38 | 35 | 0 | 3 | 0 | $0.0206 | verify raw | |
| 20261103-060647 | mistralai/mistral-large | 38 | 19 | 1 | 18 | 0 | $0.1488 | verify raw | |
| 20261103-120118 | mistralai/mistral-large | 38 | 20 | 1 | 17 | 0 | $0.1576 | verify raw | |
| 20261103-141525 | anthropic/claude-opus-5.5 | 38 | 33 | 0 | 5 | 0 | $0.2051 | verify raw | |
| 20261103-141847 | anthropic/claude-haiku-5.5 | 38 | 31 | 0 | 7 | 0 | $0.0093 | verify raw | |
| 20261103-142119 | google/gemini-3.8-flash | 38 | 32 | 0 | 6 | 0 | $0.1026 | verify raw | |
| 20261103-142737 | deepseek/deepseek-v4.1-flash | 38 | 34 | 0 | 4 | 0 | $0.0189 | verify raw | |
| 20261103-144754 | x-ai/grok-4.20 | 38 | 28 | 1 | 9 | 0 | $0.0312 | verify raw | |
| 20261103-145056 | mistralai/mistral-large | 38 | 16 | 3 | 19 | 0 | $0.1505 | verify raw | |
| 16 campaigns | — | 608 | 392 | 6 | 210 | 0 | $1.3741 | — |
Counts sum: 392 + 6 + 210 + 0 = 608. Spend is read from each campaign's append-only budget ledger, which must reconcile with the evidence pack's own claims or verification fails them apart.
The same 608 cases, split by the technique class each case belongs to. The interesting row is encoding: more cases ended inconclusive than passed, and none failed. A class the analyser mostly cannot adjudicate is a finding about the analyser, and it is on the site for the same reason the inconclusives are.
Bars are passed, inconclusive, failed, left to right, as a share of each class. Read from the per-case verdicts in the published evidence packs.
Roughly a third of all published cases resolved to neither pass nor fail. The dominant population is judge-unclear: leak probes answered with truthful deflections, and benign-appearing compliance. The conservative judge leaves both unresolved by design, because a judge that resolves ambiguity in its own favour is not a judge.
Smaller populations are hedged compliance — engaged and disclaimed in the same response — and empty responses, where there is simply nothing to judge.
Several subjects appear more than once. Published packs are immutable because receipts attest them, so when the analyser improves, the honest fix is not to mutate an old pack — it is to re-run the campaign live and publish a new chihna alongside the old one.
Both rows stay in the table. The earlier campaign is not deleted and its inconclusive count is not quietly revised.
The receipt profile is published as a standalone normative document: a third party can implement a compatible emitter or verifier from it alone, without reading the reference implementation's source. The protocol is donated the way ACME was donated.
The receipt format, its required fields, the honest-statistics rule, and the receipt-tier verification algorithm.
Prepared as an IETF individual draft in RFCXML v3 and idnits-clean. It has not been submitted; media-type registration is gated on submission.
The dependency-free receipt-tier verifier, running in the browser and in node. It is the same file this site serves on the verifier page.
The runtime half is deliberately replaceable. Execution, signed and compliance receipts can carry the actions; runtime.ledger_head_sha256 is the join point to the withstood half.
Every receipt linked above contains digests and counts only — no prompts, no responses, no keys, no credentials. The evidence artifacts stay with the issuer; the receipt is what travels.