Sep 13 14:07 Use the paper title and document benchmark contents and release links Sep 5 05:44 fig1_paper.py: legend inset moved and split over two lines 05:44 Fig. 2: keep the bubble-size legend inside its box 05:44 Fig. 2: keep the bubble-size legend inside its box 05:42 Add the figure/table build scripts and the fig data.json they read Sep 4 20:44 Model card: point at results/final and the paper names 20:44 Add the final results: tables, figures, cost panel, significance dump, results.xlsx 06:12 Revised cost: overview + zoomed benchmark figure, shared paper style 06:12 Revised figs: overview + zoomed benchmark figure, shared paper style 06:08 Final batch 2026-09-04: corrected cost panel (225/225 ok), figures, significance, error dumps, cost provenance Sep 3 05:26 Source dumps and full provenance: manifests, LR selections, cost panel, all logs 05:06 Deliverables: combined results (33/36 cells), per-run metrics from 111 runs, revision review 04:22 Final: identity pair + 3 transformer seeds close all deferred significance pairs Sep 2 20:13 Exp 7/8 operator study: fa_local, fa_sm, and the 3-D/4-D asa/fa backfill Sep 1 15:38 Correct the status line and record the cohort fix The sweep is complete (477 declared runs, 697 checkpoints); the card still said the N-D arms were partial. Also records the survivorship leak in cohort selection that these checkpoints were retrained after, with its measured effect. Aug 29 18:00 Evaluation results: leaderboards, rolling folds, subgroup fairness, cost panel, error dump, significance output, paper tables 17:55 Checkpoints: 697 runs (complete matrix) 05:47 Checkpoints: 694 runs (sweep in progress) 05:43 Checkpoints: 684 runs (sweep in progress) Aug 28 16:23 Drop checkpoints with no completion record: nothing binds those bytes to a declaration 16:21 Checkpoints: 349 runs (sweep in progress) 16:13 Checkpoints: 352 runs (sweep in progress) 09:13 Trained checkpoints: Exp 0 LR search, Tests 1/1.1 complete, N-D arms partial 05:38 initial commit 
