Back to Documentation

Claim Coverage

A precise explanation of what coverage measures and how it is determined.

TL;DR

  • Coverage measures support within the sources you supply.
  • High coverage = strong support found; low coverage = no support found.
  • Requirements map coverage to allow / deny / needs_review decisions.

Definitions (FAQ)

What is a claim?

A single, testable assertion expressed in plain text. Claims are decomposed into minimal assertions (subject, relation, attributes) for verification.

What is coverage?

Coverage is an explicit, deterministic numeric measure of how many claims are supported by the sources you provide. It is computed per the canonical definition below and returned together with per-claim evidence (quote spans and provenance).

Coverage = (number of claims with at least one supporting evidence passage whose score ≥ threshold) / (total number of claims)

  • Coverage ∈ [0.0, 1.0]
  • Deterministic — no stochastic or LLM-based inference influences the value
  • Based only on supplied sources (no web retrieval, no external inference)
  • Each claim is evaluated independently; "supported" means evidence_count > 0 AND best_score ≥ threshold

Supported · Unsupported · Ambiguous

  • Supported: A matching passage or strong paraphrase exists in your sources.
  • Unsupported: No supporting passage is found for the assertion.
  • Ambiguous: Weak, conflicting, or partial signals require caution or review.

How coverage is calculated (conceptual)

1. Identify checkable parts

Verifact identifies the individual parts of a claim that can be checked against your sources. This makes it possible to evaluate each part independently.

2. Match against sources

For each part, Verifact looks for passages in your sources that provide support. When a supporting passage is found, Verifact records the excerpt and its provenance (which source and where it appears).

3. Combine results

Verifact combines the per-part outcomes into an overall coverage score that indicates how much of the claim is supported by your sources. If all parts are supported, coverage will be high; if only some parts are supported, coverage will be partial; if none are supported, coverage will be low or zero.

Interpreting coverage

Coverage is a signal about how much of a claim is supported in the sources you supplied. It helps you decide whether to proceed automatically, ask for review, or gather more evidence.

  • High coverage: Strong, direct support is found for most parts of the claim — typically safe to proceed automatically (subject to requirements).
  • Medium coverage: Mixed or partial support — some parts are supported while others are ambiguous or missing; consider manual review or conservative enforcement.
  • Low / zero coverage: Little or no support found in the supplied sources — do not proceed automatically; further investigation is needed.

Note: contradictory passages or missing evidence reduce coverage. Endpoints return per-claim statuses and reasons that explain which parts failed or were ambiguous; use those details to guide enforcement or human review.

Concrete examples

High coverage

Claim: "Acme allows sharing usage data with subprocessors."

Source excerpt: "We may share usage data with subprocessors to provide and improve the Services."

What Verifact returns: High coverage; claim verification response includes the supporting quote and source provenance.

Partial coverage

Claim: "Acme retains user data ≤ 30 days; IPs are not kept beyond 30 days."

Source excerpt: "We retain aggregated metrics for up to 30 days. Certain logs, including IP addresses, may be retained up to 90 days."

What Verifact returns: Medium coverage; verification response highlights which parts are supported and which are contradicted, with quote spans and provenance.

Unsupported

Claim: "Acme exposes an unauthenticated public endpoint that deletes accounts."

Source excerpt: "Deletion procedures describe a support-driven workflow but no public unauthenticated endpoint."

What Verifact returns: Low/zero coverage; claim verification response contains no supporting quotes and indicates missing support.

Common pitfalls

  • Providing sources that don’t contain the needed passage — coverage will be low even if the claim is true elsewhere.
  • Claims that combine multiple facts — combined claims are harder to fully cover; split them when possible.
  • Time‑bound claims without dates in sources — include dated evidence for time-scoped assertions.
  • Using coverage directly as authorization — coverage is a signal; use requirements to map it to allow/review/deny.

How coverage is used

Coverage is a compact, technical input for authorization decisions. Authorization uses per-claim support to decide whether an action is allowed, denied, or routed for review — see the Authorization docs for details.