Buckets:
| # NUMBERS POLICY — CWM3 ICRA/IROS submission | |
| Every number in the paper must be traceable to one of the VERIFIED SOURCES below, with the file and | |
| line/section cited in a LaTeX comment next to it (or generated by `tables/gen_tables.py`, which records | |
| the source path in the emitted `.tex`). When a hand-written summary (`docs/decisions.md`) disagrees with | |
| a results table, the TABLE wins. | |
| ## VERIFIED SOURCES (numbers may ONLY be taken from these; cite the file and line/section for every number you use) | |
| - `/home/USER/cwm3/docs/progress_2026-09-11.md` (verified progress report §0–§12.11; 285 claims cross-checked; §11 = consolidated baseline comparison tables; §12 = tables landed 2026-09-12/13) | |
| - `/home/USER/cwm3/docs/results_cwm3.md`, `/home/USER/cwm3/docs/results_cwm3_hashes.md` (final results + frozen run hashes) | |
| - `/home/USER/cwm3/docs/model_spec.md`, `/home/USER/cwm3/docs/model_stage_report.md` (architecture, parameter counts, training/selection) | |
| - `/home/USER/cwm3/results/tables/*.md` (BATON tables: stage_final, stage_final_5seed, stage_final_ablations, stage_I1_v2b, k4_v2_summary, exp3_P, warning_head_final*, warning_head_final_oof, operating_points, power_gold, robustness_final, kill_report_final, branch_arch_ablation, stage_F, stage_F_rssm, stage_final_gzfc, stage_final_ncroi, gaze_chain, identity_invariance, state_supervised_E, c1_transfer, latent_*) | |
| - `/home/USER/cwm3/aide/results/tables/*.md` and `/home/USER/cwm3/aide/docs/protocol.md` (AIDE: stage_a, stage_bc (incl. §1b/§1c published rows + margins, §8 kill table, §9 5-seed), stage_a_5seed, stage_bc_5seed, tradeoff_all_vs_hm, analytic_baselines, baton_pretrain); published Driver-WM rows in `/home/USER/cwm3/aide/eval/paper_rows.py` | |
| - `/home/USER/cwm3/docs/decisions.md` (pre-registration rows and verdict rules; hand-written summaries — when it disagrees with a table, the TABLE wins) | |
| - `/home/USER/cwm3/docs/newtest_prereg.md`, `/home/USER/cwm3/docs/newtest_prereg_deviations.md`, `/home/USER/cwm3/docs/annotation_programme.md` (human-evaluation programme: pre-registered design; labels are still being collected — report as in progress, no results) | |
| - Paper skeleton, thesis, reviewer-attack list and fact checks: `<local-notes>/plans/elegant-dazzling-stallman.md` (§六, §七, and supplement B1) | |
| ## VERDICT WORDING RULES | |
| (kill_checks.beats: ≥2/3 seeds CI excl. 0, same sign, |mean Δ| > seed sd): BEATS → "significantly better"; WORSE → "significantly worse"; INCONCLUSIVE/NOT_WORSE → "matches / no significant difference" — never "beats". Pre-registered FAIL stays FAIL (B3, B4, B5 "matches, does not exceed", B1 4/6). Never convert enriched false-alarm shares into natural rates. Never merge the two AIDE model rows (5.8 M geometry vs 6.3 M vis+sem) into one model. Appendix variants (gaze chain, non-causal ROI, 5-seed) never replace headline numbers. Negative results are reported as such. | |
| ## CITATION RULE | |
| only add a bibliography entry when you are certain of title, authors, venue and year; otherwise write `\cite{TODO_shortname}` and list it in `/home/USER/cwm3/paper/TODO_citations.md`. Never invent references. | |
| ## FILE RULE | |
| write ONLY the file(s) assigned to you; do not edit other sections; use `\input{tables/<name>}` for the agreed generated tables instead of typing numbers when a table exists. | |
Xet Storage Details
- Size:
- 3.34 kB
- Xet hash:
- 6bd0ca4627b909f80aad7e7c8383a85094635a7f7f498480b3af1bae4c7a4b9a
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.