Anchored Whitening Layer / report_bench_2026-09-02T234052.md
Failed on benchmark
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Built a matched cnn_small vision comparison on the CIFAR-10 subset: the baseline was standard channel-mixing CNN behavior, while the idea inserted an EMA ZCA whitening layer after the first convolution. Across 8 paired seeds, the idea had higher test error (0.870625 vs 0.853750), with paired delta +0.016875 and permutation p=0.50765; therefore there is no significant win. The behavior-based signature also failed to confirm the claimed mechanism: observed off-diagonal covariance remained 1.9904, minimum fidelity was -0.0010 versus rho_min=0.8, and unit-variance error was 0.99997.", "metrics": { "baseline": "vision/cnn_small; baseline sweep lr={0.001,0.003,0.01}; best lr=0.003; full 8-seed test error mean=0.853750, std=0.032476; per_seed=[0.890000,0.840000,0.870000,0.785000,0.870000,0.850000,0.890000,0.835000]", "idea": "EMA Q=I anchored whitening after conv1; shared lr grid; selected lr=0.003; full 8-seed test error mean=0.870625, std=0.046465; per_seed=[0.760000,0.855000,0.885000,0.910000,0.855000,0.910000,0.900000,0.890000]", "paired_comparison": "delta_mean=+0.016875 (idea-baseline; lower is better), idea_wins=2/8, permutation_p=0.50765, verdict=no significant win", "mechanism_signature": { "prediction": "A trained anchored whitening layer should reduce channel covariance while preserving designated-channel fidelity.", "idea_offdiag_cov_observed": 1.9903746843338013, "idea_max_unit_variance_error_observed": 0.9999727606773376, "idea_min_fidelity_observed": -0.0010492571163922548, "rho_min": 0.8, "confirmed": false } }, "bench_report": { "bench_version": 1, "track": "vision", "model": "cnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.003 }, "sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.8537499755620956 }, { "cfg": { "lr": 0.003 }, "mean": 0.846249982714653 }, { "cfg": { "lr": 0.01 }, "mean": 0.8699999898672104 } ], "full": { "mean": 0.8537499830126762, "std": 0.032475961531880884, "per_seed": [ 0.8899999856948853, 0.8399999737739563, 0.8700000047683716, 0.7849999666213989, 0.8700000047683716, 0.8499999642372131, 0.8899999856948853, 0.8349999785423279 ], "n": 8 } }, "idea": { "mean": 0.8706249967217445, "std": 0.046464878782190286, "per_seed": [ 0.7599999904632568, 0.8550000190734863, 0.8849999904632568, 0.9100000262260437, 0.8549999594688416, 0.9100000262260437, 0.8999999761581421, 0.8899999856948853 ], "n": 8 }, "comparison": { "delta_mean": 0.0168750137090683, "idea_wins": 2, "n_pairs": 8, "per_seed_diffs": [ -0.12999999523162842, 0.01500004529953003, 0.014999985694885254, 0.12500005960464478, -0.01500004529953003, 0.060000061988830566, 0.009999990463256836, 0.05500000715255737 ], "p_value": 0.50765, "mde": 0.061319939059687474, "mde_rel_pct": 7.182423458833265, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "prediction": "trained anchored layer should reduce channel covariance while preserving designated-channel fidelity", "baseline_offdiag_cov": null, "idea_offdiag_cov_observed": 1.9903746843338013, "idea_max_unit_variance_error_observed": 0.9999727606773376, "idea_min_fidelity_observed": -0.0010492571163922548, "rho_min": 0.8, "confirmed": false }, "idea_sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.8587500005960464 }, { "cfg": { "lr": 0.003 }, "mean": 0.8525000065565109 }, { "cfg": { "lr": 0.01 }, "mean": 0.885000005364418 } ], "protocol_note": "shared lr grid; 8 paired seeds; 400/200 CIFAR subset; 4 epochs" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py > bench_results.json", "files": [ "bench_experiment.py", "bench_results.json" ], "limitations": "The benchmark used the standard 400/200 CIFAR-10 subset and 4 epochs rather than the README's example 30 epochs, due to runtime constraints. The implementation used Q=I and did not optimize an orthogonal Q or enforce the rho_min constraint with a penalty; the signature probe was collected from one trained idea model and no baseline covariance signature was recorded. The observed signature failure may therefore reflect both the simplified implementation and short training, but it does not provide evidence for a win.", "system_verdict": "failed", "practical_verdict": "inconclusive", "mechanism_ok": 0, "system_judged": true }