Skip to content
VeraFrame Docs

The trust report

The trust report is VeraFrame’s primary output. It accompanies every document, chatbot audit, answer, or JSON response produced by the platform.

The report describes the truth of the generated content, not its style. It never rewrites what the AI produced. It tells you whether each factual claim can be supported by your own source material, and if not, what is wrong.

The overall confidence band

At the top of every report there is a single confidence indicator:

  • High — most claims verified against the source material, no mismatches
  • Medium — some claims could not be verified, or low-severity issues present
  • Low — at least one mismatch or fabricated claim, or a large fraction of claims unverified

The band is a heuristic summary. It is intended to route your attention. Do not use it as a publish/block signal by itself — the per-claim indicators below are what you act on.

Per-claim indicators

For each factual claim in the output, the report assigns one of four statuses:

Verified

The claim matches a specific passage in your source material. The report shows:

  • The direct source text
  • The normalized value (for numbers and dates)
  • The file and location the match was found in

A verified claim is safe to publish as-is, subject to your own editorial review.

Needs review

The claim could not be found in the source material — not because it is wrong, but because VeraFrame did not find evidence for it. The report shows:

  • What claim was made
  • Which sources were searched
  • Why no match was produced (for example, the term was not present, or a numeric range did not overlap)

A “needs review” claim is not the same as a false claim. It may be true but unsupported by the provided sources. Your job is to either accept it, add missing source material, or remove the claim.

It is also not automatically the same thing as a workflow review case. In tenants where review workflow is disabled, needs review stays a report-level signal only. In workflow-enabled tenants, the same finding can be routed into the review queue if the tenant’s review settings say so.

Mismatch / Fabricated

The claim conflicts with your source material, or it makes a specific factual assertion that no source supports. The report shows:

  • The claim that was made
  • The source passage that contradicts it (for mismatches)
  • A “fabricated” marker (for claims with no plausible source)

A mismatch is a concrete correction signal. Do not publish without resolving it.

Missing

A field or required element that your request or template expected is absent from the output because it was also absent from the source material. The report shows:

  • The field that was expected
  • Where in the output it should have appeared

Missing claims distinguish “the AI did not produce this” from “the AI produced the wrong thing.”

Source conflicts

When two source documents disagree on the same fact — for example, two contracts with different payment terms — VeraFrame surfaces the conflict rather than silently choosing one. The report shows both sources and asks you to pick the correct one before final output.

Math and schema checks

For numeric content, VeraFrame also runs arithmetic and unit checks (for example, that a sum of line items matches a total) and, for JSON outputs, validates the output against the schema you provided. These checks appear as dedicated sections in the report.

Acting on the report