Pass rates, from real requests
Every number here comes from a request somebody actually made and paid for. Nothing is sampled, curated or rerun to look better. A model judged fewer than 20 times shows its count and withholds its rate.
By model
No verdicts yet. The board fills itself from the first judged request.
How to read it
A pass is what STANDS after a human has had their say. The judge's own rate is beside it, and the gap between the two is how often the machine was wrong — which is why the overrule count is in the same row.
The judge is advisory everywhere except the text gate: an output that was asked for specific words and did not render them is a hard fail and cannot be talked round. An output nothing could read is counted as unread and never as a pass, so a board cannot climb by the judge going quiet.
Read from the ledger at 2026-09-19 04:45:06Z.