Parity-block curvature preconditioner / report_bench_2026-09-03T122941.md

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP

Стенд-проверка (stage-2) · промт оператора:

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 10, "verdict": "The parity-block AdamW intervention was tested on the registered tabular track with the same branch-swap two-branch MLP as the baseline. It performed significantly worse: baseline test MSE was 0.2395 versus 0.3616 for the idea, with paired delta +0.1221 and p=0.0081. The mechanism signature observed separated even and odd gradient energy, but no task-metric win occurred.", "metrics": { "baseline": "Best lr=0.01; full 8-seed test MSE mean 0.2395111453.", "idea": "Best tested lr=0.003; full 8-seed test MSE mean 0.3616223410.", "delta_mean": "+0.1221111957; p=0.0081; idea wins 0/8." }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "two_branch_mlp", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01 }, "sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.3566400483250618 }, { "cfg": { "lr": 0.003 }, "mean": 0.27394459396600723 }, { "cfg": { "lr": 0.01 }, "mean": 0.23046058416366577 } ], "full": { "mean": 0.23951114527881145, "std": 0.04028734767167621, "per_seed": [ 0.1893874704837799, 0.2472962588071823, 0.2523687779903412, 0.23278982937335968, 0.1838403344154358, 0.25920143723487854, 0.3209288716316223, 0.2302761822938919 ], "n": 8 } }, "idea": { "mean": 0.3616223409771919, "std": 0.09797382371900515, "per_seed": [ 0.28693538904190063, 0.2831035852432251, 0.28846070170402527, 0.3903844356536865, 0.3120247721672058, 0.30676016211509705, 0.5750100016593933, 0.4502996802330017 ], "n": 8, "idea_sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.545, "n": 8 }, { "cfg": { "lr": 0.003 }, "mean": 0.3616223409771919, "n": 8 }, { "cfg": { "lr": 0.01 }, "mean": 0.489 } ] }, "comparison": { "delta_mean": 0.12211119569838047, "idea_wins": 0, "n_pairs": 8, "per_seed_diffs": [ 0.09754791855812073, 0.035807326436042786, 0.03609192371368408, 0.15759460628032684, 0.12818443775177002, 0.047558724880218506, 0.254081130027771, 0.2200234979391098 ], "p_value": 0.0081, "mde": 0.0701771215456511, "mde_rel_pct": 29.300148627304555, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "prediction": "branch-swap parity separates even and odd gradient energy", "trained_model_gradient_norms_plus": 0.0739268725739474, "trained_model_gradient_norms_minus": 0.05150565393582205, "ratio_minus_over_plus": 0.6967108460270031, "confirmed": true }, "track_justification": "Optimizer intervention; the official bench assigns optimizer ideas to tabular, and both systems use the same branch-swap symmetric MLP and differ only in the update rule." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 official_stage2.py", "files": [ "official_stage2.py", "bench_report.json", "official_stdout.txt" ], "limitations": "Only the registered tabular track was tested. The experiment used a small two-branch MLP and did not test vision, sequence, dynamics, wall-clock speed, FLOPs, or empirical sector-specific stability-boundary scans.", "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }