Attestsearchesfortheevidenceandtellsyouwhetheraclaimistrue.Attesttellsyouwhetheraclaimholdsupagainstyourowndocuments.Attest tells you whether a claim holds up against your own documents.

ev_1docs/api.md, Spans

supportingSpans and contradictingSpans are span ids in sourceId:ordinal form, where the ordinal indexes sentences within that evidence item. ev_1:0 is the first sentence of evidence item ev_1. Every returned span id resolves to text you sent. Ids the model invents are stripped before the response is built, and the fraction that survived is reported in audit.referentialIntegrity.

ev_2Developer landing, Attest

Attest takes a claim and the evidence behind it and returns a verdict with the exact sentences that support or contradict it. It never searches for evidence and never rules on the world, only on the documents you supply, and nothing you send is kept.

ev_1Developer landing, rules

SUPPORTED means supported by what you sent. Attest never retrieves. If the evidence does not cover the claim you get INSUFFICIENT_EVIDENCE, and Attest cannot tell whether no such evidence exists or your retrieval missed it. Nothing is stored. Attest holds claims, evidence and uploaded files in memory for the request. The gateway logs the route, status, latency and claim count of each call, and nothing else.

ev_2Developer landing, Attest

Attest takes a claim and the evidence behind it and returns a verdict with the exact sentences that support or contradict it. It never searches for evidence and never rules on the world, only on the documents you supply, and nothing you send is kept.

ev_1docs/api.md, MCP server

Attest is also exposed as an MCP tool, on the same code path as the HTTP API. The server identifies itself as Quorum Attest, with the Quorum mark and a link to quorumtech.ch, so it is recognisable in a client that lists several servers side by side.

ev_1docs/api.md, errors

400: the request does not match the schema. error.issues names the failing field. 502: the upstream model provider failed. 503: Attest could not measure (INDETERMINATE), or no provider is configured. INDETERMINATE returns 503, not 200. It means Attest failed to measure, which is a server-side failure rather than a finding about your claim. Returning it with a 200 would let it read as a verdict.

ev_1src/attest/service.ts, comment on AttestBase

There is no confidence score. Measured over 24 real checks, 22 landed between 0.95 and 0.99 and the only other value was a constant standing in for a flag, so the field carried no information while looking as though it did. It was removed.

ev_1docs/privacy.md, the rule

Attest never persists files, evidence or claims. That is a design property of the product, not a limitation of the prototype, and every layer below is built so that it stays true without anyone having to remember it. A caller sends Attest a claim and the documents behind it. Attest reads them, measures them, answers, and forgets. Nothing that arrives in a request is written to disk, to a database, to a log, or to any store that outlives the process. Uploaded files are parsed from the request buffer into process memory and live until the batch is cleared, evicted, or the process restarts. Claims and evidence sent inline live in process memory for the duration of the request. The verdict, spans and provenance live in the response, and nowhere else. The container enforces it: the Attest image runs on a read-only root filesystem with all capabilities dropped, so a bug that tried to spill a board pack to disk fails rather than succeeding quietly.

ev_1docs/privacy.md, the gateway

Authentication, limits and the dashboard live in the enterprise gateway, which sits in front of Attest and is the only thing that reaches it. The gateway does have a database, and it is deliberately blind to content. Its collections can hold an API key registry with id, name, tenant, scopes, limits and timestamps, never the key itself. OAuth clients, grants and SHA-256 hashes of tokens. Usage buckets per credential, per UTC hour, per route: request count, error count, rate-limited count, claim count, total latency. A request log with one row per request with its time, credential, route, status, latency and claim count, kept 30 days. There is no field in any of those for a claim, a document, a span or an evidence id, and the code path that records usage reads only the request's shape: which route, how many claims. Request bodies are streamed through the gateway to Attest and are never buffered whole, logged, or written anywhere.

ev_1docs/privacy.md, what a customer can rely on

The only durable trace of a document is the SHA-256 fileHash Attest returns per upload, so an audit can later ask whether this verdict was produced against the document I signed off by hashing the file it still holds. A hash is not the document and cannot be turned back into one. A verdict is reproducible from the document they still hold plus the hashes and versions in the provenance block. It is not retrievable from us, because we do not have it. Revoking a key or disconnecting a client leaves behind only counts. Restarting Attest leaves nothing behind at all.

ev_1docs/api.md, Accuracy

Measured on 190 hand-labelled examples covering multi-hop reasoning, arithmetic, distractors, quantifier and scope traps, unit prefixes, near-miss paraphrase and conflicting sources. Verdict accuracy 95.9%. Overclaim rate 1.6%. Grounding-degree accuracy 96.2%. Refutation accuracy 97.9%. Source-conflict accuracy 98.9%. Overclaim rate is the number to care about: how often Attest asserts more grounding than the evidence supports. It is the failure that costs you something. Two caveats worth stating plainly. Run-to-run variation is about 3.7 points at temperature zero, because hosted inference is not bit-reproducible, so treat small differences as noise. And these figures come from one labelled set built by one author; they describe this distribution, not yours.

ev_1docs/measurement-notes.md, order invariance

The voting run doubles as an order-sensitivity test, and this is the one genuinely new finding: unanimous across three rotations 179 of 190, 94.2%. Order changed the verdict on 11 of 190, 5.8%. Evidence ordering flips the verdict on 5.8% of cases, and order carries no information about grounding, so every one of those is a robustness defect rather than a borderline judgement. The remaining errors are not mostly cases the model cannot do; they are cases where it is unstable, and which way it lands depends on span order.

ev_1docs/measurement-notes.md, scoreboard

Scoreboard. Baseline, single call: 92.6% flat accuracy, 93.0% grounded, still unbeaten. Two-family replicas: 80.0%, about 20x wall clock. Hybrid, lean policy: 74.4%. Stance replicas: 74.6%. Nothing has beaten a single grounding call. The decomposed architecture has now lost under every configuration tried: field-name rubric, sentence rubric, holistic hybrid, model-family replicas, and stance replicas. Three remaining levers were tested against the baseline: reasoning effort high was 90.5%, worse; self-consistency voting was 93.6%, no gain, and on the 11 split votes the majority was correct only 4 times. Nothing tried has beaten a single well-prompted grounding call.

ev_1Gateway tariff, version 2, effective 2026-09-05

One credit ($0.01) for a claim checked against the evidence in the request, two ($0.02) for one checked against a loaded batch, explained or not. Per claim in the array: one credit ($0.01) with evidence in the request, two ($0.02) against a loaded batch. A draft review reserves maxClaims (default 200, at most 500) at two credits each and settles to two credits ($0.02) per claim the review checked.

ev_2Developer landing, pricing

One credit is one cent, forever: prices change by changing rows, never the value of a credit. Credits per claim: one checked against the evidence in the request, two against a loaded batch.

ev_1docs/api.md, Accuracy

Measured on 190 hand-labelled examples covering multi-hop reasoning, arithmetic, distractors, quantifier and scope traps, unit prefixes, near-miss paraphrase and conflicting sources. Verdict accuracy 95.9%. Overclaim rate 1.6%. Grounding-degree accuracy 96.2%. Refutation accuracy 97.9%. Source-conflict accuracy 98.9%. Overclaim rate is the number to care about: how often Attest asserts more grounding than the evidence supports. It is the failure that costs you something. Two caveats worth stating plainly. Run-to-run variation is about 3.7 points at temperature zero, because hosted inference is not bit-reproducible, so treat small differences as noise. And these figures come from one labelled set built by one author; they describe this distribution, not yours.

ev_1Developer landing, rules

SUPPORTED means supported by what you sent. Attest never retrieves. If the evidence does not cover the claim you get INSUFFICIENT_EVIDENCE, and Attest cannot tell whether no such evidence exists or your retrieval missed it. Nothing is stored. Attest holds claims, evidence and uploaded files in memory for the request. The gateway logs the route, status, latency and claim count of each call, and nothing else.