-
Fri, Nov 6, 2026 40
-
Stage G: final build of the revised manuscript (82 pages)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Stage F B1-B2 N1-N10 (+P3b)
…
Stage F corrections after the verifier's sign-off NO (HEAD 77f38ec). B1: H1 reported as the pre-stated seed-aware rule returns it - supported in 2024 at a small magnitude (4/4 horizons on scaled CRPS, every seed pair negative, 1.5-3.4 x the seed half-band, 3.0-4.5 x the MDE), "within a seed half-band" / "bounded null" removed, budget qualifier kept; generated verdict sentence paper/tables/h1_verdict_sentence.tex. P3b: the eight 12k seed checkpoints scored once on the untouched 2025 origins (M3U96b + M4M64c, batch_N1j/N5b/N1k), seed_contrast --tag final -> Table 10 Panel D, Tables B.10-B.13, *Final* macros: the 2024 verdict is not replicated by the rule in 2025 (t_seed -1.9..-2.3) although 4 of 5 pairs keep the sign. B2: abstract/S1/S11 quote the 12k model on scaled CRPS (1.54 %/0.30 %, Table 8; GainRwtSmallScaledTest*). N1: pre-registered wording of H1-H5 restored as the primary statement with "operationalisation adopted in revision" and a S9 deviations paragraph. N2: subgroups on both CRPS conventions (pairwise.py, recomputed on M3U96b), S7.4 quotes the scaled values. N3/N4: abstract 147 words with both H2 clauses. N5: generated plural, DM-inflation file set named. N6: venue-not-API-verified comments in references.bib (generator flag). N7: nOfflineTests dropped, rel-QLIKE macros suppressed when rel_valid is false, gallery/exh build artefacts and *_sel.npy ignored, main_legacy/*.json committed, PHASE_7 revision note. N9: economics note names the diagnostic rows. Build: 82 pages, 0 errors, 0 undefined, 0 overfull, 0 MacroMissing. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Stage E exhibits (Agent 9), final pass: seeds-vs-HAR pairwise family recomputed on M3U96b against the corrected HAR-RV (unresolved item 16-iii; table note updated), main_mcs_heldout re-input in Appendix B after the underscore fix (item 16-i), compound surnames braced in build_references.py and references.bib regenerated (69/69 re-resolved), copula figure panel titles, exhibit map/notes, final build paper/main.pdf (78 pages, 0 errors, 0 overfull, 0 undefined, 0 MacroMissing)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Stage E writing (Agent 10), pass 3: coverage macros in percent (baseline fix sentence), signed contrasts worded as contrasts (§1, §7.2, §7.6, §11)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 final run on the complete N batch (3d1c3f7): corrected-baseline tables, recalibration before/after tables, economics/dir_acc/recal/package records in inline_numbers.json, Figures 1-5 regenerated (WP41-EDIT-01, WP41-EDIT-02, WP41-ADVE-02, WP41-ADVE-11, WP41-AI_C-11, WP41-DATA-01, WP41-ECON-06, WP41-ECON-07, WP41-THEO-01, WP41-ECON-04, WP41-EDIT-04, WP41-ADVE-08, WP41-DATA-11, WP41-ECON-11, WP41-DATA-03, WP41-DATA-13)
…
Final numbers (2024, scaled CRPS unless stated): MRPT-S vs corrected HAR-RV at H = 5 -0.07 % [-0.13, -0.01], t_cl -2.24, Holm 0.15 (was -1.82 %, t -58); recalibrated twins: MRPT-S_recal vs HAR_recal -0.40 % (t_cl -14.6), vs ridge_recal -0.13 % (Holm 0.005); corrected LightGBM 5.4 % worse than the random walk at H = 5 (the only cell below it; other 47 cells 2.89-4.87 % unscaled, 2.50-4.14 % scaled); directional accuracy ridge 0.5181 [0.5152, 0.5209] vs neural 0.4998-0.5007 at H = 5; economics with the delay: 5 of 224 positive cells in 2024 (11 expected false positives), 0 in 2025 (AR-41-1 not triggered); archive 171 files / 27.6 GB; compute ledger 37 runs, 12.1 GPU-h (6.0 baseline study). 104 tables, 14,069 + 297 + 133 macros; latexmk 44 p., 0 errors, 0 overfull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P5: recalibrated pairwise contrasts (2024, 2025) and remaining summaries/families
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P11: QA-flag count restricted to the 300 universe instruments (242 flagged rows: confirms the §4.4 sentence)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P11 results (M3U96b): restricted-release archive manifest (171 files, 27.6 GB, sha256 of the tar), archive reproduction check, QA-flag recount (universe + look-ahead instruments), environment record, FinData provenance records (API-call logs, no bars)
…
Items: WP41-DATA-03, WP41-DATA-06, WP41-DATA-10, WP41-DATA-14, WP41-THEO-17 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P6 results (M3U96b): H5 economics on seven location models with the 1-minute execution delay (280 cells x 4 tags, summaries with cell counts and expected false positives; AR-41-1 not triggered), directional accuracy with day-cluster CIs; P5 recalibrated held-out leaderboards; P7 copula-head loadings
…
Items: WP41-EDIT-04, WP41-ADVE-08, WP41-DATA-11, WP41-ECON-11, WP41-THEO-02, WP41-THEO-11, WP41-AI_C-08 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P6: directional accuracy undefined (NaN) for zero-location models; economics re-run chain
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P11: per-node compute ledger tables (compute_table.csv, compute_runs.csv) from the gathered ledger and training summaries (WP41-DATA-06)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P5 (acceptance 8): pre-fix baselines re-run on m4mn with the legacy construction — cross-node tolerance vs the committed M3U96b leaderboard
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P1 final pass + P5 corrected baselines (m4mn refit, M3U96b pass): main/final pairwise (60/72-test Holm families incl. naive_rw_t and transfer300 pairs), MCS, stress contrasts, transfer-gap intervals, compare tables incl. crps_scaled, leaderboards with the corrected HAR/ridge/EN/LightGBM rows, recal/{c_val,nu_hat,lightgbm_floor,baseline_fix_delta}, mde.csv, convention_check.csv, node ledger
…
Items: WP41-EDIT-02, WP41-EDIT-11, WP41-EDIT-13, WP41-ADVE-01, WP41-ADVE-07, WP41-ADVE-09, WP41-ADVE-12, WP41-AI_C-05, WP41-AI_C-13, WP41-AI_C-17, WP41-DATA-04, WP41-DATA-08, WP41-ECON-01, WP41-ECON-04, WP41-ECON-05, WP41-ECON-06, WP41-ECON-08, WP41-ECON-09, WP41-ECON-10, WP41-ECON-12, WP41-ECON-15, WP41-THEO-01, WP41-THEO-06, WP41-THEO-12, WP41-THEO-13, WP41-THEO-14, WP41-THEO-17, WP41-EDIT-03, WP41-EDIT-05, WP41-AI_C-06 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P2 results (m4mm): 40 checkpoints re-scored on the original 3,840-row single-instrument subset and on the 50,000-origin stratified subset — no 12k model changes selection; 8 of 15 4k ablations would select last.pt on the stratified subset
…
Items: WP41-ADVE-05, WP41-AI_C-01, WP41-DATA-05 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 post-pull run 6 (P5 recal nu_hat / c_val / lightgbm_floor, P10 ablations_h3 pairwise + stress contrasts + decomposition + MCS): Table 12 note from lightgbm_floor.csv, ablations_h3 tables, H3-control and decomposition records in inline_numbers.json (WP41-THEO-05, WP41-AI_C-04, WP41-AI_C-15, WP41-THEO-13, WP41-ECON-05, WP41-THEO-14)
…
H3 control (reduced global token, stats vs none, stress days): the stats variant has LARGER |coverage gaps| at 50/80 % with day-cluster CIs excluding zero at H = 5/15/30 (dabs_cov_50 +0.49/+0.36/+0.31 pp) and a wider PIT variance -> 0 of 4 horizons closer to nominal, 3 of 4 farther (H3 not supported; CRPS: stats -0.07 % at H = 60, t_cl -6.1). Decomposition vs the 2023-calibrated Student-t random walk: time-only share 61/48/31/-36 % at H = 5/15/30/60 (85-89 % against the Gaussian RW). LightGBM floor: 0.053 % of H = 5 origins below 0.01 sigma^2 H, raw QLIKE -4.47 -> -11.83 floored. 100 tables, 13,193 macros; latexmk 44 p., 0 errors, 0 overfull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P10: time-only decomposition against naive_rw_t and naive_rw from the committed pairwise files (WP41-THEO-05, WP41-AI_C-15)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P10 results (M3U96a redglob twins, M5M36 time_only_redglob + shuffled_pv, contrasts on M3U96b): H3 control without the global-token confound, stress-day paired coverage-gap contrasts, time-only decomposition against naive_rw_t, price/volume-only shuffle
…
Items: WP41-THEO-05, WP41-AI_C-04, WP41-AI_C-15, WP41-THEO-13 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
gather_node.sh: exclude only the copied naive_rw reference files (naive_rw_t arrays must be gathered)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N1 tail chains trimmed to the remaining steps (P3/P4/pass 2 done in N1g/N1h; rescore_validation on m4mm)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 post-pull run 5 (P4 learning curve 12k/24k/36k + budget verdict, ablations_h3 complete): budget tables, learning-curve and H3-control records in inline_numbers.json, compute ledger rows (WP41-EDIT-08, WP41-AI_C-02, WP41-THEO-09, WP41-THEO-05, WP41-AI_C-04)
…
budget_verdict.json (P4): criterion (i) fails at H = 5 and 15 (the 36k H1 gap +0.05 % / -0.01 % falls outside the five-seed band of the 12k contrast) and holds at H = 30 and 60 (-0.09 / -0.25 %); criterion (ii) holds (MRPT-S vs HAR-RV keeps its sign at 30/60 min, +0.49 / +0.62 % at 36k vs +0.53 / +0.65 % at 12k) -> the budget condition stays in the abstract. Validation-NLL slope per 1,000 steps at 36k: -0.00001 (MRPT-S, converged) vs +0.00006 (1m-only). latexmk 42 p., 0 errors, 0 overfull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P4 results (M3U96a + M5M36 chains, curve on M3U96b): learning curve 12k/24k/36k of the Small twins with budget_verdict.json (H1 gap at 36k leaves the seed band at H=5/15: budget condition stays); P2 results: copula head retrained with the stratified validation (identical selection, step 400 -> identical joint/indep scores; Table 9 unchanged), ablations pass 2; P9 transfer300 training record and held-out summaries
…
Items: WP41-EDIT-08, WP41-AI_C-02, WP41-THEO-09, WP41-ADVE-05, WP41-AI_C-01, WP41-THEO-03, WP41-EDIT-05, WP41-AI_C-06 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 post-pull run 4 (P8 lookahead66 2024/2025 leaderboards + pairwise, ablations_h3 pairwise, recal neural c): second held-out set tables (both splits, stems lookahead66 / lookahead66_final), lookahead66 block and macros in inline_numbers.json (WP41-EDIT-06, WP41-ADVE-03, WP41-THEO-04, WP41-EDIT-05, WP41-AI_C-14)
…
lookahead66 (66 availability-excluded instruments, zero-shot checkpoints, baselines refit on their histories): 63,182 origins (2024) / 83,322 (2025); every neural model beats the random walk by 4.5-5.2 % (scaled, 2024) and 4.6-5.0 % (2025); LightGBM is 3.2 % (2024) and 4.2 % (2025) WORSE than the random walk at H = 5 on these names; MRPT-S vs HAR-RV -0.82 % at H = 5 (t_cl -10.8) and +0.08/+0.33/+0.36 % at 15/30/60 (Holm 0.80/0.08/0.11) in 2024, -0.85 % then -0.04/+0.09/+0.17 % (Holm >= 0.75) in 2025; H1 on the 66: -0.07..-0.15 % (2024, Holm 0.008-0.13), -0.03..-0.00 % (2025, Holm 1). 91 tables, 11,677 macros; latexmk 42 p., 0 errors, 0 overfull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P8 results: second held-out set of 66 availability-excluded instruments — baselines refit on their own histories (M3U96b), zero-shot predictions of the five main checkpoints on 2024 and 2025 (M2U64), leaderboards, MCS, pairwise with cluster-t; P5 neural recalibration constants fitted on the 100k 2023 subset (M2U64)
…
Items: WP41-EDIT-06, WP41-ADVE-03, WP41-AI_C-07, WP41-THEO-04, WP41-DATA-14, WP41-THEO-01 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 post-pull run 3 (P3 seed_contrast.csv / seed_bands.csv): Figure 5 with the five-seed band of MRPT-S 12k, seed contrast records in inline_numbers.json (WP41-EDIT-07, WP41-AI_C-03, WP41-AI_C-16, WP41-ADVE-13)
…
seeds/seed_contrast.csv (P3): H1 scaled CRPS 2024 detectable at every H under the two-test rule (t_seed -4.53/-3.17/-3.19/-4.05 with 4 df, t_cl on the seed-averaged differential -13.6/-14.0/-13.0/-11.2); unscaled detectable at H = 5 and 60 only (t_seed -3.54/-2.47/-2.33/-2.83). Seed band of MRPT-S scaled CRPS: half-band 0.055-0.062 %; of the 1m-only twin 0.066-0.137 %. latexmk 42 p., 0 errors, 0 overfull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 post-pull run 2 (seeds, budget and ablations_h3 partial, lookahead66 baselines): compute ledger with the node/chip of every review run (32 runs, 10.3 GPU-h incl. the 6.0 of the baseline study), seeds tables, seed band and per-seed H1 pairs in inline_numbers.json (WP41-EDIT-07, WP41-ADVE-13, WP41-AI_C-03, WP41-AI_C-16, WP41-DATA-13)
…
Per-seed H1 (seed k full vs seed k 1m-only, scaled CRPS, 2024): negative in all five seeds at every H; mean -0.081/-0.117/-0.152/-0.214 % (min..max -0.146..-0.039 / -0.260..-0.047 / -0.336..-0.065 / -0.413..-0.127), t_seed (4 df) -4.53/-3.18/-3.20/-4.05; unscaled t_seed -3.54/-2.47/-2.34/-2.84. Seed half-band of MRPT-S scaled CRPS 0.055-0.062 %. The official seed-aware contrast (seeds/seed_contrast.csv, P3 on N1) supersedes these leaderboard-derived numbers when it lands. latexmk 42 p., 0 errors, 0 overfull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P3 results: seed study of the H1 twins (seeds 2-5 on M4M64a/M4M64b/M4M64c, contrast on M3U96b) — five-seed H1 contrast with t(4) and seed-averaged cluster-t, seed bands, seed-vs-HAR pairwise, leaderboard; eligibility import fix; N1b3/N1g chains
…
Items: WP41-EDIT-07, WP41-ADVE-13, WP41-AI_C-03, WP41-ECON-02 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
Offload P8 zero-shot + P5 neural validation predictions to M2U64 (N7b) and P2 checkpoint re-scoring to m4mm (N8b)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P8: origin eligibility — share of origins removed by the future-presence rule per split and role (M3U96b)
…
Items: WP41-AI_C-14 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P8: instrument status from the FinData catalog (CNO/CWH/HASI/PSEC/SHO/IRBT/GLOB; 300/300 universe grids have 2025 bars), instrument_master_version filled in universe_main.json
…
Items: WP41-DATA-09, WP41-THEO-04, WP41-DATA-14 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P8: instrument_status fix (string session dates), N1b2 chain; waitloop.sh
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 post-pull run 1 (P1 pass 1 main/final + ablations single family, P7): twin scaled tables, day-cluster t in every compare table, MCS tables, generated Gaussian note, \pw macros on scaled CRPS with t_cl/p_holm, headline numbers with t_cl ranges (WP41-EDIT-01, WP41-EDIT-02, WP41-ADVE-02, WP41-ADVE-11, WP41-ECON-06, WP41-ECON-07, WP41-AI_C-11, WP41-DATA-04, WP41-THEO-14)
…
inline_numbers.json: 2024 t_cl range vs RW -42.7..-21.5 (DM -71.6..-25.5); H1 2024 t_cl -4.40/-3.67/-3.39/-4.12 (DM -5.35/-4.54/-4.50/-5.25), Holm on p_cl in a 56-test family; H1 scaled -0.127..-0.039 % (2024), -0.005..+0.022 % (2025); Gaussian head +0.32/+0.20 % with day-cluster t 20.2 / 6.5 in the single 68-test family. 67 tables, 8,597 macros; latexmk 42 p., 0 errors, 0 overfull. Pass-1 files are overwritten by the final P1 pass (corrected baselines) -> one more make tables after that pull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
P1 (ablations, pass 1 on m4mm) + P7 (m4mm): single Holm family of 68 CRPS tests (0 NaN, Gaussian contrast included, stale *_gaussian files removed), MCS, stress contrasts; copula diagnostics on the saved PIT scores; P8 coverage audit 2017H2-2022 and look-ahead export records (M3U96b)
…
Items: WP41-ADVE-04, WP41-AI_C-08, WP41-ECON-03, WP41-THEO-03, WP41-EDIT-02, WP41-ADVE-01, WP41-ECON-01, WP41-EDIT-06 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L5: Holm (1979) JSTOR identifier kept as a note, not a DOI (doi.org and Crossref return 404 for 10.2307/4615733); jof250 audit: 0 unknown keys, 43/43 remaining DOIs resolve (WP41-LITE-08)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L4: Figure 3 (PIT histograms) regenerated on M3U96b with the short panel titles (WP41-AI_C-19)
…
Generated from a /tmp overlay of the worktree code reading ~/wp41_uqo/results through a symlink (nothing written on the data home); the laptop cannot regenerate this figure (per-origin PIT arrays). Log: reproducibility/WP-41/logs/L4_fig_pit_M3U96b.log. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L7 (docs, Makefile): REPRODUCIBILITY.md with the program-to-exhibit map, nu=inf re-score provenance, restricted archive, tolerances, stale-artefact policy, multi-node recipe; make tables / figures / test-review targets (WP41-DATA-03, WP41-DATA-06, WP41-THEO-17, WP41-AI_C-16)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
L1/L4 generators (pre-pull run on the baseline aggregates): headline numbers, convention-labelled tables, generated Table 9 / Gaussian note / MCS / compute table, Figures 1B, 3, 4B, 5 (WP41-EDIT-01, WP41-EDIT-02, WP41-ADVE-02, WP41-ADVE-11, WP41-ADVE-15, WP41-AI_C-11, WP41-DATA-01, WP41-DATA-12, WP41-ECON-06, WP41-ECON-07, WP41-LITE-07, WP41-THEO-07, WP41-EDIT-14, WP41-ADVE-14, WP41-AI_C-18, WP41-AI_C-19, WP41-AI_C-20, WP41-DATA-13, WP41-DATA-04, WP41-THEO-14, WP41-ADVE-13)
…
scripts/review/headline_numbers.py -> results/inline_numbers.json + paper/inline_macros.tex: every abstract/§1/§7 range on both conventions and all four splits (2024: 2.19-4.89 % unscaled, 1.79-4.17 % scaled; 2025: 2.20-4.42 / 1.78-4.01; held-out 2.50-4.44 / 2.05-3.93 and 2.24-4.00 / 1.90-3.78; neural-only 3.68-4.27 %; H1 -0.10..-0.04 % (2024) vs +0.01..+0.03 % (2025); HAR gap 0.24-0.50 % scaled; gap to EWMA-t -0.03..0.64 %, to seasonal 0.33-1.11 %; DM -72 to -26; pilot: LightGBM worse than RW at every H). scripts/evaluate.py: notes block on every table naming sample, inference and convention; twin <exp>_compare_crps_scaled.tex (written once P1's compare_crps_scaled files are pulled); t_cl used when present; QLIKE in levels when rel_valid is false; WIS x 10^4; parameter column; \pw macros with t_cl/p_holm/p_cl/lo/hi when present (convention auto: crps_scaled when available); Table 9 generated from path_scores json; Gaussian note generated from the single-family pairwise file; MCS tables; compute ledger table. Figures: 1B names the convention, 3 short panel titles, 4B relabelled (3,840 index-ordered origins) with re-score overlay hook, 5 as differences vs mrpt_small_b4k with the seed band (order-filter bug in the forest fixed). Appendix: \FloatBarrier so Appendix B tables stay under their heading and the ticker list is contiguous; Figure 4 referenced in §5.5; compute table under Table 1. latexmk: 39 p., 0 errors, 0 overfull. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-
N1 tail chains (recalibration merge/re-scoring, economics, loadings, validation re-score; P10/P3/P4 contrasts, pass 2, final P1 pass; P11 archive)
…
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
-