…L_PREFIX in the #82 scorer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ontrols' two missing splits

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… rows, doc

Three control fine-tunes (released weights + 2,000 GSV steps, seeds 1-3) and the
released checkpoint scored on all 14 reviewed splits with the #82 scorer; the #82
controls topped up on bayonne/juneau as a same-seed rerun read. control_arm_263.py
reads the committed caches (seed-mean and per-seed paired pano bootstrap, seed
spread, unbiased subset, Laurens rig effect, determinism check). Compute ledger:
11 klone rows, 37.29 GPU-h, $0, from a committed sacct -D dump.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… rule, training logs

M1 0.55 seed mean shown with each seed's interval (seed 1 is above released);
M2 rig-gap widening attributed to the GoPro arm, GSV arm flat at max-F1;
M3 same-seed rerun listed with its dataset/commit/GPU/resume confounds;
M4 unmeasured cross-GPU bound removed; M5 the six training logs committed
under analysis_out/control_arm_263/trainlogs/ in #82's layout; M6 reading
rule marked proposed. N7-N10 wording and tests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>