Skip to content
VeraFrame Docs

Reading the trust report

This page is a practical companion to The trust report concept. Here, we walk through what you see in the UI and how to act on it.

The overall confidence badge

Every trust report opens with a single badge at the top:

  • High (teal background)
  • Medium (amber background)
  • Low (red background)

Next to the badge is the issue count — the number of claims that are flagged as needs-review, mismatch, or missing.

Use the badge as routing, not as a decision. Low confidence means you must look at the details. High confidence means the details will still matter for regulated outputs.

Source validation

The largest section of the report is Source validation. It lists every factual claim the validator tried to verify and its status:

  • A green checkmark for verified claims — click to expand and see the direct quote.
  • An amber indicator for needs review — no matching source passage found.
  • A red indicator for mismatch / fabricated — the claim conflicts with the source, or is unsupported.

Each row shows:

  • The claim as it appears in the output.
  • The source passage (if any) that was matched against it.
  • A normalized value (for numbers, dates, and percentages — so “€1,200.00”, “1200 EUR”, and “1.2k€” are compared as the same value).
  • The source file and location.

Math validation

For any numeric claims that involve arithmetic — “the total is €X”, “the sum of line items equals €Y” — the report includes a Math validation section. It shows:

  • The expression that was computed.
  • The expected result.
  • Whether the result in the output matches.

A math mismatch means the narrative and the numbers disagree with each other. It is a common failure mode for AI-generated financial or quote content.

Completeness

The Completeness section confirms that every required field (from a template schema, an output schema, or the implicit structure of your request) was actually produced. Missing required fields appear here.

Source conflicts

When two source documents disagree on the same fact, a dedicated Source conflicts section surfaces the disagreement. For example, if two contract versions specify different payment terms, both versions are shown and the field is flagged for manual selection.

Resolve source conflicts before publishing. Choosing silently is how an AI produces confident-but-wrong output at scale.

Evidence panel

Clicking any finding opens the Evidence panel on the right side of the screen. It shows:

  • The output passage in context.
  • The source passage in its full surrounding text, not just the matched snippet.
  • The retrieval audit — which other source blocks were considered and why they were not chosen.

The retrieval audit is what makes a VeraFrame output defensible later. It is exported as part of the audit trail for regulated deployments.

Acting on the report

For each flagged finding you have three options:

  • Approve — the claim is fine, even though the validator could not directly verify it. Approvals are recorded.
  • Override — the validator flagged a mismatch but the source is wrong or incomplete. Overrides are recorded with your note.
  • Edit — rewrite the output passage and rerun validation.

See Correcting findings for details.