CVaR-tail active residual correction / report_bench_2026-09-02T010752.md

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP

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

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 9, "verdict": "Implemented a matched tabular benchmark using the standard mlp_tiny architecture and an upper-tail residual-correction training intervention, with shared learning-rate sweep, paired eight-seed evaluation, and a trained-model mechanism signature. The mechanism prediction was confirmed: tail-to-global MSE ratio fell from 1.964 to 1.142 and CVaR absolute error fell from 2.056 to 0.858. However, standard test MSE worsened from 7.4623 to 7.8551, with paired permutation p=0.0235 and verdict "idea worse (significant)", so this does not count as a win.", "metrics": { "baseline": "tabular/mlp_tiny test MSE mean 7.4622584581, std 0.6335098399; best lr=0.01; sweep means for lr 0.001/0.003/0.01 were 16.9858301/11.5920208/7.1914142.", "idea": "tabular/mlp_tiny with global-MSE warmup followed by tail-weighted residual correction, best lr=0.01 and tail_weight=4.0: test MSE mean 7.8551361561, std 0.8610502822; paired delta +0.3928776979; permutation p=0.0235; tail/global MSE ratio 1.1424 versus baseline 1.9637; CVaR absolute error 0.8576 versus baseline 2.0555." }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "weight_decay": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0 }, "mean": 16.985830068588257 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 11.592020750045776 }, { "cfg": { "lr": 0.01, "weight_decay": 0.0 }, "mean": 7.1914142370224 } ], "full": { "mean": 7.462258458137512, "std": 0.633509839885441, "per_seed": [ 6.857808589935303, 7.703956604003906, 7.029021263122559, 7.174870491027832, 6.467283725738525, 7.960057258605957, 8.130547523498535, 8.37452220916748 ], "n": 8 } }, "idea": { "mean": 7.855136156082153, "std": 0.8610502821731341, "per_seed": [ 7.060885906219482, 8.29471206665039, 7.688426494598389, 7.102557182312012, 6.474584579467773, 9.063955307006836, 8.353429794311523, 8.80253791809082 ], "n": 8 }, "comparison": { "delta_mean": 0.3928776979446411, "idea_wins": 1, "n_pairs": 8, "per_seed_diffs": [ 0.2030773162841797, 0.5907554626464844, 0.6594052314758301, -0.07231330871582031, 0.007300853729248047, 1.103898048400879, 0.22288227081298828, 0.42801570892333984 ], "p_value": 0.0235, "mde": 0.3233736709994596, "mde_rel_pct": 4.333455787058997, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "idea_config": { "lr": 0.01, "tail_weight": 4.0 }, "idea_sweep": [ { "cfg": { "lr": 0.001, "tail_weight": 4.0 }, "result": { "mean": 22.177693843841553, "std": 0.9958955479883927, "per_seed": [ 22.019962310791016, 20.866891860961914, 22.155040740966797, 23.668880462646484 ], "n": 4 } }, { "cfg": { "lr": 0.003, "tail_weight": 4.0 }, "result": { "mean": 11.351011991500854, "std": 0.7376803433578396, "per_seed": [ 10.290882110595703, 11.659903526306152, 12.307673454284668, 11.145588874816895 ], "n": 4 } }, { "cfg": { "lr": 0.01, "tail_weight": 4.0 }, "result": { "mean": 7.536645412445068, "std": 0.5031108845030456, "per_seed": [ 7.060885906219482, 8.29471206665039, 7.688426494598389, 7.102557182312012 ], "n": 4 } } ], "mechanism_signature": { "prediction": "tail residual correction should preferentially reduce upper-tail prediction error relative to global error", "baseline_behavior": { "global_mse": 7.462258815765381, "tail_mse": 14.653417646884918, "tail_to_global": 1.9636705196993307, "cvar_abs_error": 2.055535450577737 }, "idea_behavior": { "global_mse": 7.855136334896088, "tail_mse": 8.973360180854797, "tail_to_global": 1.1423557527564545, "cvar_abs_error": 0.8575644314289099 }, "predicted_tail_focus": true, "observed_tail_to_global_ratio_change": -0.8213147669428762, "confirmed": true } }, "protocol_notes": { "epochs": 24, "batch": 128, "baseline_grid": [ { "lr": 0.001, "weight_decay": 0.0 }, { "lr": 0.003, "weight_decay": 0.0 }, { "lr": 0.01, "weight_decay": 0.0 } ], "idea_grid": [ { "lr": 0.001, "tail_weight": 4.0 }, { "lr": 0.003, "tail_weight": 4.0 }, { "lr": 0.01, "tail_weight": 4.0 } ], "intervention": "same mlp_tiny trained with global-MSE warmup followed by upper-tail weighted residual correction" } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py", "files": [ "stage2_bench.py", "bench_report.json" ], "limitations": "Only the structurally matched built-in tabular regression track was tested. The benchmark intervention is a supervised tail-weighted residual-loss approximation rather than an actual expensive-label acquisition loop or neural ensemble; no vision, sequence, dynamics, or downstream constrained optimization experiment was run.", "system_verdict": "failed", "practical_verdict": "harms", "mechanism_ok": 0, "system_judged": true }