Explanation, audit and provenance

A verdict alone is a decision. The rest of the result is the reasoning, the check on that reasoning, and the record that lets someone reproduce it later.

Explanation

For a claim checked against inline evidence, three rubric axes mark propositions about the claim, each with its reasoning and the spans it relied on:

{
  "explanation": {
    "axes": {
      "entailment": {
        "covers_every_assertion": {
          "reasoning": "Only the revenue assertion is supported; the operating margin improvement lacks evidence.",
          "holds": "false",
          "spans": ["ev_1:0"]
        }
      }
    },
    "citedSpans": ["ev_1:0"],
    "availableSpans": 2,
    "axesRun": ["entailment", "contradiction", "scope"],
    "axesExpected": ["entailment", "contradiction", "scope"]
  }
}

holds is "true", "false" or "unclear". The explanation does not decide the verdict. It is produced independently and can disagree with it.

For a claim checked against a batch, the explanation is different in shape: reasoning, the list of unverifiedAssertions (the actionable output), searchedThoroughly, and a trace of what was searched and read, in order, with the number of steps taken.

Audit

{
  "audit": {
    "referentialIntegrity": 1,
    "explanationTrust": 1,
    "findings": [],
    "explanationDisagrees": false
  }
}

• referentialIntegrity is the share of cited spans that resolve to text you sent, from 0 to 1. A citation the model invented was removed and lowers this.
• explanationTrust is that figure less mechanical coherence penalties.
• findings names problems, for example fabricated_span_reference.
• explanationDisagrees: true means the explanation read the evidence differently from the verdict. It is not an error. It flags a borderline case and is surfaced rather than reconciled away. In practice it fires on DISPUTED cases, where the contradiction axis sees a refutation while the verdict declines to pick a side.

Provenance

{
  "provenance": {
    "measurementHash": "3f9c...",
    "policyHash": "a41b...",
    "rulesApplied": ["precedence", "adjacent-not-partial"],
    "promptVersion": "attest-routed-v1+precedence+adjacent-not-partial",
    "scoringVersion": "attest-scoring-v1",
    "serviceVersion": "attest-service-v0.1",
    "model": "openai/gpt-oss-120b",
    "createdAt": "2026-09-06T14:02:11.318Z",
    "mode": "evidence"
  }
}

• measurementHash identifies the inputs: the canonical claim, the evidence and the measurements. policyHash identifies the scoring configuration. They are separate so that changing a threshold does not invalidate a measurement no model call would produce differently.
• rulesApplied lists the reading rules chosen for this request. The set is part of promptVersion, so two responses with the same version were produced by the same instructions.
• mode is evidence or batch-agent.

A verdict is reproducible from the document you still hold plus these hashes and versions. It is not retrievable from Quorum, because Quorum does not keep it. See Privacy.

Degraded

{ "degraded": ["explanation_incomplete: 2/3 axes"] }

Non-fatal problems, so you always learn when an answer was produced with reduced measurement. An empty array means nothing was compromised.

Switching the explanation off

config.explain: false skips the explanation pass. You get the verdict at a fraction of the latency, but explanation and audit come back null and the span lists are empty, so there is no audit trail and no citations. It costs the same credits as an explained check, on purpose, so nobody is nudged away from the audit trail. Reach for it when you are triaging volume, not when the answer matters.

Keeping an audit trail

Store, per claim you checked: the claim, the evidence ids and their fileHash or content, the verdict, the span ids, and the whole provenance block. That is enough to answer "was this verdict produced against the document I signed off" by hashing the file you still hold, and to replay the check against the same versions.