Evidence and review
Read source grounding, meaning, and confidence as separate parts of a result.
Evidence policy 13: production rollout
Production rollout verified on 24 September 2026: the hosted API includes evidence policy 13 and the hosted MCP connector runs 1.9.0. This does not publish or upgrade the npm package installed by local clients. Inspect response metadata and the installed client version; saved responses retain their original policy and evidence time.
Provider-generated summaries and snippets are discovery material. A redirect resolving to a publisher is not proof that its page was read. A failed read remains unavailable evidence, not a contradiction or a finding that the claim is false. Legacy records without receipts remain unknown; this update does not rewrite historical records.
What a source match establishes
grounded reports whether a quotation was bound back to its source. It does not establish that the source itself is correct or that the quote preserves the original meaning.
Binding methods are exact, normalized, fuzzy, reflowed, and unbound. Fuzzy and reflowed matches stay needs_review. Keep matched_text so a reviewer can inspect the passage.
Read the verification stamp
layer:deterministic,bound,model, orunbound, describing the evidence basis.band:verified,needs_review, orunverified, describing the review state.confidence: a number or null. A score is not a guarantee.grounded: whether source text was matched.meaning:not_checked,faithful,changed, orambiguous. Missing assessment must not be presented as faithful.review_reason: why the result needs review, when available.
Use the whole stamp. Filtering only on a high score can hide a meaning problem or an unresolved match.
OCR readings and provenance
A prepared OCR reading may remove page furniture or join lines. Preserve both the original matched source text and the derived reading. Keep transformations, droppedLines, retainedSpans, calibrationKey, basis, model, and decidedAt.
A reflowed quote remains needs_review even when the derived text matches exactly. Prepare OCR evidence requires retained source and representation IDs. See the batch reference for a complete illustrative OCR result.
Read supported corrections
When evidence establishes a replacement for a wrong clause, verdict.corrections can contain a corrected_statement and exact proof quotes linked to citations from the same claim group. correction_status: not_established means no replacement was proven. Do not invent a corrected value when none is returned.
Read a publisher evidence receipt
In policy 13, a citation can carry an optional evidence object with version: 1. It records a particular publisher read, not an independent certification of the publisher's claims.
| Field | Meaning |
|---|---|
originalUrl, finalUrl | Discovered URL and captured destination; a null destination was not established. |
state | discovered, matched, read, unavailable, unsupported, or invalid. matched identifies an exact selected passage from extracted publisher text; read does not establish a relevant match. |
contentHash, retrievedAt | SHA-256 identity and capture timestamp, or null if unavailable. A new fetch is a new capture. |
extractionVersion | publisher-blocks-v6 for new captures; earlier extraction versions remain valid for historical receipts. |
quoteMatch | exact, normalized, none, or not_attempted. URL resolution and HTTP 200 alone do not establish an exact match. |
retention, replayable | Current receipts use hash_only and false. A hash does not retain the source bytes needed to replay the historical page. |
failureReason | Why evidence could not be used, or null. |
capture | Optional HTTP status, content type, extracted-text hash, and robotsNoarchive observation. |
context | Optional list of up to four separate adjacent publisher passages, at most 500 characters each. These provide context and are never joined into the selected quote. |
dates | Attributed publisher metadata: value, precision (day or year), kind (published or updated), and basis: publisher_metadata. An empty list means no accepted metadata date. |
Receipt matching is distinct from batch quote binding, which supports additional matching methods. Keep the receipt with the exact citation and claim it describes. Publication and update dates do not establish when a legal or institutional change became effective. A year in a forecast, URL, copyright notice or generated summary is not a publication date.
Separate source scores from verdict confidence
credibility_score: null means unknown; 0 is a known zero. credibility_basis distinguishes heuristic, measured, and unknown. Policy 5 labels retained source-ranking estimates as heuristic. Provider grounding confidence measures attribution confidence and must not be interpreted as source credibility. Historical measured labels alone do not establish how a score was calculated.
Verdict confidence_basis: unknown and confidence_available: false mean calibrated truth confidence is unavailable. The legacy numeric confidence is retained as aggregation_score for compatibility; calibrated_confidence is null. Do not display that aggregate as a probability that the claim is true. A finding's model-reported confidence, a source's relevance and the overall verdict confidence answer different questions; their percentages need not match. Read the cited passages, unresolved timing, conflicting evidence and uncovered claim scope. Unknown authority is not proof that a source is false, and an official source is not automatically decisive for every claim. For current administrative affiliation only, a matched direct official statement can take precedence over undated sources with unverified authority. The summary explains that choice and conflict retains the opposing quotation, even when the result is supported or contradicted. The source is not labelled stale without evidence. Conflicting official records, dated counter-evidence and fact-checks remain relevant when they concern the same period. A dated report of a past event does not by itself assert that its organizational affiliation still holds today; explicit continuity claims and retrospective challenges remain counter-evidence.
Search provider and historical limits
The current development repair keeps provider-derived discovery separate from publisher evidence. Alternate search provider implementations are not yet supported; a configuration switch alone cannot enable Tavily, Serper or Firecrawl. Unavailable or unsupported evidence remains explicit.
Derived response and search caches are versioned and isolate execution scope. A cache hit does not refresh the publisher capture time. Old entries and missing historical provenance cannot be relabeled as current evidence.