Reviewing a flagged answer

Opening a row shows the request in detail: request id, model, artifact and time, Flagged and New badges, and key numbers (sequence risk 0.91, threshold 0.50, 92% probability unreliable, 10 tokens out).

Risk per token

The answer panel shows a token-by-token risk strip, the score after each token, a chart of how the reader escalated and peaked, and where Latent's read sits in calibration.

Record a verdict

A side panel lets the reviewer Confirm failure or mark a False alarm (with counts), run Analyze with judge and pick suggested flag tags (wrong number, invented entity, contradicts the source…), with a note field. "Why it was flagged" explains the decision in plain words.

Similar Details screens

All Details

More from Latent

All Latent screens