| 14:30 |
0–1 |
Project created after WP22; protocol (truncated) completed; WP22 inputs inspected; data builders (Task A horizon-dated features, Task B vintage panel), evaluation module (DM/GW/MCS/bootstrap) |
| 14:45 |
2–3 |
Models A (trees + torch) and B (classic + trees + torch); OpenMP conflict → process families; HSHN fixed (embedding init, horizon training); runs launched on M3U96b |
| 15:00 |
4 |
Simulation design (5 DGPs on the WP22 simulator); 100 synthetic panels built on the node; literature agent: 152 verified references + novelty scan |
| 15:30–16:30 |
5–6 |
Task A evaluated; ablations (36/37), interpretability (SHAP/ALE), robustness (38) run; sections written as results arrived |
| 17:00–17:40 |
6–7 |
Task B evaluated; simulation forecasters summarised (10/6 replications); all sections written; 44-page manuscript compiles with 0 missing macros; anonymous version; pushed to wp23_uqo.git |
| 19:00–20:30 |
7 (v1.1) |
Author review: appendices had bare headings and an empty « Supplementary tables » section → rewritten (architecture equations + TikZ diagrams, text before every table, test-method appendix with Holm/GW tables, cost table, FloatBarrier). New scripts/39_extra_analysis.py on the fetched Task A/B predictions: errors by origin/year and by ex-post regime, Holm-adjusted DM, directional accuracy and turning points, P(decline) Brier/reliability, equal-weight combinations, municipal MASE distribution, RAREF filtered regimes, dollar errors, selection rates, computational cost → 11 new tables, 5 new figures, new subsections in results / heterogeneity / probabilistic / interpretation / data. Wide tables scaled (fit). 56 pages, 0 undefined references; anonymous version rebuilt. |