Models

Ordering. Models rank by benches fully passed, then mean score relative to each board's best published model (1.00 = board leader). Per-bench chips show valid passes; the flag pill counts audited sessions that failed the reward-hack review — it is displayed, never a sort key.

Roster vs charts. This directory lists every published model, including superseded siblings the homepage charts curate away. Click any row for its per-problem cells, audit chips, GPU boards, and integrity record — or open the run index for raw transcripts.