VALIDATION TRANSPARENCY

Benchmark history.

Every capture run we have published, newest first, with the change against the run before it. Each row opens the full report for that run — same page, same methodology, the numbers as they stood on the day.

Runs

Deltas compare each run with the one below it. For word error rate, lower is better. For coverage, higher is better. Hallucination counts are per corpus, and confirmed means a fabricated claim that a majority of jurors agreed on; provisional means only one juror flagged it.