Skip to content
EverrunsEverruns

Citation Verification

IDcitation_verification
CategoryKnowledge
Featurescitations
GuardrailYes
DependenciesNone

Check that each citation is faithful — that the source it points to actually supports the sentence it is attached to — and stamp a verdict on it. It is a guardrail capability, decoupled from the feeds: it consumes the citations collected during a turn (from Retrieval Citations or any future feed) and verifies them uniformly, so any feed can be paired with any verifier.

It contributes no tools and no system prompt. The verdict renders in the UI as a badge on each source.

Verdicts

Section titled “Verdicts”

Each citation gets one of three verdicts, shown as a badge on the chip preview and in the Sources strip:

VerdictMeaning
entailedThe source supports the claim. Rendered as a green verified badge.
unsupportedThe source does not support the claim. Rendered as an amber unsupported badge.
uncertainSupport could not be established. Rendered as an unverified badge.

An assistant message whose Sources strip shows three cited sources with verdict badges: source 1 "verified" (green), source 2 "unsupported" (amber), and source 3 "unverified", produced by citation_verification over the retrieval feed's citations.

Modes

Section titled “Modes”
ModeCostBehavior
heuristic (default)FreeDeterministic lexical entailment — token overlap between the claim span and the source snippet. No model call. A weak but honest baseline, strongest on verbatim citations.
llmOne utility-model call per citationA utility-model judgement per claim/source pair (claim = hypothesis, snippet = premise). More accurate; falls back to the heuristic when no utility model is configured.

Configuration

Section titled “Configuration”
FieldTypeDefaultDescription
mode"heuristic" | "llm""heuristic"Verification strategy (see above).
thresholdnumber (0–1)0.5Entailment threshold for the heuristic verdict — the fraction of the claim’s distinctive tokens the source must cover to be entailed.

Notes

Section titled “Notes”
  • Feed-agnostic: verifies citations from any feed via the shared render contract, so evals can hold the feed fixed and vary only the verifier.
  • Enabled by default: citation_verification is part of the generic (default) harness in heuristic mode, so citations are verified out of the box with no model cost.
  • llm mode needs a utility model: with none configured it degrades gracefully to the heuristic rather than failing.
  • No new data egress: the verifier reasons only over text the feed already retrieved.

See Also

Section titled “See Also”