Run a structured fact-check before the final answer

Verification workflow for checking claims, citations, evidence boundaries, Chain-of-Verification, and confidence before using AI-generated output.

Purpose

This guide defines a final fact-checking and claim-verification workflow. It is not the evidence-boundary chooser. First select or confirm the active evidence boundary, then use this workflow to decompose the draft or answer into claims, verify each claim against the allowed evidence, check citation or artifact locators, run Chain-of-Verification, and return either a corrected final answer or a fail-closed result.

When to use this

Use this section to decide whether this workflow is the right fit before you configure prompts, policies, or reference material.

  • Use case
    A generated answer contains factual claims
    Use this when an AI-generated answer, draft, report, summary, article, or technical explanation contains claims that need evidence support before final use.
  • Use case
    The answer will be published, sent, reused, or used for a decision
    Use this when unsupported claims, weak citations, or invented evidence would create publication, product, academic, engineering, operational, legal, or reputational risk.
  • Use case
    The workflow already has an active evidence boundary
    Use this after selecting files-only, cited authoritative sources, web verification, or academic evidence mode. The fact-checking workflow verifies within that boundary; it does not replace the boundary decision.
  • Use case
    The output needs a final confidence signal
    Use this when a non-sentinel final answer should include an evidence-based confidence score after claim support, citation quality, and uncertainty have been checked.

Step-by-step implementation procedure

Follow the workflow in order. Each step gives one action and one verification check before continuing.

  1. Step 1 · Confirm the active evidence boundary

    Identify whether the answer must be checked against provided artifacts, cited authoritative sources, web-verified sources, or academic evidence mode.
    If no evidence boundary is selected, stop and route to the evidence-boundary decision guide before fact-checking.
    Check: The review states the active evidence boundary before evaluating claims.

  2. Step 2 · Separate the draft into checkable claims

    Extract material factual claims, technical claims, source claims, quantitative claims, claims about standards or documentation, and claims that affect the conclusion.
    Create a claim list and exclude purely stylistic or subjective wording unless it implies a factual claim.
    Check: Each material claim can be matched to evidence, corrected, removed, or marked unsupported.

  3. Step 3 · Map each claim to allowed evidence

    Check every material claim against the source type allowed by the active boundary.
    Use artifact locators for files-only claims, stable citations for public-source claims, and reviewed-source locators for academic evidence mode.
    Check: No claim is treated as verified because it sounds plausible or because it appeared in memory or prior chat context.

  4. Step 4 · Check citation and locator quality

    Verify that citations, file references, section references, source names, and locators actually support the claims they are attached to.
    Flag weak, vague, mismatched, stale, inaccessible, or non-authoritative citations according to the active policy.
    Check: Every retained claim has a locator that supports the exact claim, not just the general topic.

  5. Step 5 · Run Chain-of-Verification

    Run a final second-pass check for unsupported claims, contradictions, overstatement, missing evidence, and boundary violations.
    Correct supported wording, remove unsupported claims, or return the fail-closed result required by the active policy.
    Check: The final answer is aligned with the evidence actually available.

  6. Step 6 · Apply confidence reporting only when allowed

    Add confidence only after verification passes and only when the active policy does not require a sentinel-only fail-closed output.
    If evidence is insufficient, do not append a confidence score to a sentinel-only response. If the answer is supported, report confidence based on evidence quality, coverage, and uncertainty.
    Check: Confidence does not replace citations, artifact locators, or missing-evidence handling.

  7. Step 7 · Return the final verified output

    Return a corrected final answer, a claim-level correction report, or the fail-closed result required by the active evidence boundary.
    Keep the output format aligned to the user’s requested deliverable and the active policy.
    Check: The final answer does not include unsupported claims that were identified during verification.

Verification checklist

Use this checklist before accepting the output, publishing it, or using it as evidence for a downstream workflow.

  • Boundary
    The active evidence boundary is named
    The review identifies whether claims are checked against provided artifacts, cited authoritative sources, web-verified sources, or academic evidence mode.
  • Claims
    Material claims were extracted
    The workflow identifies claims that affect factual accuracy, technical correctness, source support, or the final conclusion.
  • Evidence
    Each retained claim has allowed evidence
    Every retained factual claim is supported by an allowed artifact locator, cited source, reviewed source, or other evidence permitted by the active policy.
  • Citations
    Citations and locators support the exact claim
    The cited source or artifact locator supports the claim attached to it, not merely the surrounding topic.
  • Fail closed
    Unsupported claims stop or change the output
    Unsupported claims are removed, corrected, marked as insufficient evidence, or handled with the active fail-closed sentinel.
  • Confidence
    Confidence is evidence-based and non-sentinel only
    A confidence score is added only after verification passes and never appended to a sentinel-only fail-closed response.

Next steps