Reverb docs

Judge Validation

Why the report includes a trust badge instead of treating LLM judgment as unquestioned truth.

1

Each answer is judged against a small fixed schema: target mention, prominence, competitors mentioned, and cited domains.

2

Brand and product aliases are normalized per scorecard before trusting mention counts.

3

The validation queue samples about 40 answer judgments, or all answers when fewer are available.

4

A human reviewer labels target mention, prominence, and competitor mentions for the sampled answers.

5

The public report shows validation pending, trusted, or needs review based on field-level agreement.

6

The default trust threshold is 85% overall agreement across labeled fields.