Characteristic-Region Gain Controller / report_bench_2026-08-31T113007.md

Failed on benchmark

Raw ⬇ ZIP

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

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 9, "verdict": "Built the Characteristic-Region Gain Controller as a local training-loop intervention on the bench dynamics track using the shared rnn_small GRU. The idea achieved lower mean test MSE than the tuned baseline (0.0010705 vs 0.0014901; delta -0.0004197; 6/8 wins), but the paired permutation p-value was 0.1834, so the required significant-win criterion was not met. The model-derived feedback-radius versus local hidden-Jacobian signature had a 70.9% relative gap and was not confirmed.", "metrics": { "baseline": "Tuned Adam with lr=0.006 and gradient clipping=1.0: mean dynamics test MSE 0.0014901440, std 0.0004653271, 8 paired seeds.", "idea": "Characteristic-region controller with lr=0.006 and delta=0.10: mean dynamics test MSE 0.0010704752, std 0.0004834065, 8 paired seeds; paired delta -0.0004196688; 6/8 wins; permutation p=0.1834; verdict=no significant win." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_run.py", "files": [ "bench_run.py", "bench_report.json" ], "limitations": "Only the built-in dynamics track was tested; no sequence, tabular, or vision transfer was evaluated. The finite-delay fractional-memory characteristic equation was approximated here by a GRU recurrent-weight block spectral-radius proxy rather than an exact fractional-memory augmentation. The local Jacobian signature was measured on four trained models and did not confirm the predicted quantitative correspondence. Training used 12 epochs and 400/100 samples per seed rather than larger-scale or longer-horizon evaluation.", "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006, "clip": 1.0 }, "sweep": [ { "cfg": { "lr": 0.0015, "clip": null }, "mean": 0.0054811002 }, { "cfg": { "lr": 0.0015, "clip": 1.0 }, "mean": 0.0042505179 }, { "cfg": { "lr": 0.003, "clip": null }, "mean": 0.0039752387 }, { "cfg": { "lr": 0.003, "clip": 1.0 }, "mean": 0.0035102504 }, { "cfg": { "lr": 0.006, "clip": null }, "mean": 0.0021467398 }, { "cfg": { "lr": 0.006, "clip": 1.0 }, "mean": 0.0013746844 } ], "full": { "mean": 0.001490144, "std": 0.0004653271, "per_seed": [ 0.0009367615, 0.0008422294, 0.0017605363, 0.0019592105, 0.0009294281, 0.0016490562, 0.0019374506, 0.0019064794 ], "n": 8 }, "idea_grid": [ { "cfg": { "lr": 0.0015, "delta": 0.03 }, "mean": 0.00196241 }, { "cfg": { "lr": 0.003, "delta": 0.05 }, "mean": 0.001255191 }, { "cfg": { "lr": 0.006, "delta": 0.1 }, "mean": 0.0011763981 } ], "idea_best_cfg": { "lr": 0.006, "delta": 0.1 } }, "idea": { "mean": 0.0010704752, "std": 0.0004834065, "per_seed": [ 0.0018825745, 0.0005507631, 0.0014166068, 0.0008556481, 0.0014776132, 0.0004715494, 0.0012880854, 0.0006209611 ], "n": 8 }, "comparison": { "delta_mean": -0.0004196688, "idea_wins": 6, "n_pairs": 8, "p_value": 0.1834, "mde": 0.0006828097, "mde_rel_pct": 45.8217, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "predicted_feedback_radius": 0.4187634, "observed_local_hidden_jacobian_gain": 0.7154711, "relative_gap": 0.7085332, "n_models": 4, "confirmed": false } }, "system_verdict": "failed", "practical_verdict": "inconclusive", "mechanism_ok": 0, "system_judged": true }