# Ablation leaderboard — stage6 mode: thermometer

_completed 1/1 · failed 0 · regenerated after every row_

| rank | ablation | bpt_gap | top1_agree | student_bpt | teacher_bpt | ARC-E | HSwag | corpus | knob delta |
|---|---|---|---|---|---|---|---|---|---|
| 1 | SH | 0.5944 | 0.7316 | 3.3186 | 2.7242 | 0.730 | 0.695 | wikitext | merge_heal_enabled=True, merge_heal_train_router=True |

Ranking: lower `bpt_gap` = less compression damage. On the `nemotron` corpus the gap is RELATIVE (Stage-2.5 adaptation is common-mode across rows) — rank rows against each other, not against zero. Tiebreak: higher `top1_agree`, then ARC-E+HSwag.
