== ModelBrain (4 threads), p50 / p95 ms per repeat
locomo   locomo_r1.jsonl    top-5 differs from locked on 0 of 2304 rows
locomo   locomo_r2.jsonl    top-5 differs from locked on 0 of 2304 rows
locomo   locomo_r3.jsonl    top-5 differs from locked on 0 of 2304 rows
locomo   A_base_eub12     p50 [14.9, 14.9, 14.9] p95 [18.4, 18.4, 18.4] repeatable=True
locomo   B_rr20_eub12     p50 [173.8, 174.0, 173.6] p95 [203.7, 203.8, 204.2] repeatable=True
beam100  beam100_r1.jsonl   top-5 differs from locked on 0 of 280 rows
beam100  beam100_r2.jsonl   top-5 differs from locked on 0 of 280 rows
beam100  beam100_r3.jsonl   top-5 differs from locked on 0 of 280 rows
beam100  A_base_eub12     p50 [19.5, 19.4, 19.4] p95 [29.0, 28.9, 29.0] repeatable=True
beam100  B_rr20_eub12     p50 [556.8, 550.7, 551.5] p95 [646.1, 643.6, 644.9] repeatable=True
beam500  beam500_r1.jsonl   top-5 differs from locked on 0 of 560 rows
beam500  A_base_eub12     p50 [21.5] p95 [31.1] repeatable=single run
beam500  B_rr20_eub12     p50 [536.8] p95 [645.0] repeatable=single run