Covariance-Conditioned Neural Rollouts / report_bench_2026-09-02T160812.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented a matched dynamics-track comparison using the shared rnn_small GRU backbone: deterministic MSE baseline versus an end-to-end Gaussian behavior head trained with Gaussian NLL. The Schur-complement sanity check was positive semidefinite, and the trained Gaussian model had 95% interval coverage of 1.00 on the measured seed, confirming an uncertainty-calibration signal within the declared tolerance. However, across 8 paired seeds the idea had higher test MSE (8.59209e-05 versus 3.97188e-05), with a significant paired loss (delta_mean=4.62020e-05, p=0.03865), so it did not win.", "metrics": { "baseline": "dynamics/rnn_small; best lr=0.006; mean test MSE=3.971884348175081e-05; std=7.221693826148878e-06", "idea": "GaussianGRU with the same GRU backbone; best lr=0.001; mean test MSE=8.592086487624329e-05; std=7.508073923840228e-05", "paired_comparison": "delta_mean=4.6202021394492476e-05, p_value=0.03865, idea_wins=1/8, verdict=idea worse (significant)", "mechanism_signature": "seed-0 trained-model Gaussian NLL=-2.5754659175872803, 95% interval coverage=1.0, mean predicted variance=0.0008512148051522672, confirmed=true" }, "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006 }, "sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.0007079141651047394 }, { "cfg": { "lr": 0.003 }, "mean": 0.00019564990725484677 }, { "cfg": { "lr": 0.006 }, "mean": 4.179111101620947e-05 } ], "full": { "mean": 3.971884348175081e-05, "std": 7.221693826148878e-06, "per_seed": [ 4.624741632142104e-05, 5.4139207350090146e-05, 2.856133687600959e-05, 3.82164835173171e-05, 4.160348544246517e-05, 3.6566088965628296e-05, 3.5268141800770536e-05, 3.714858758030459e-05 ], "n": 8 } }, "idea": { "best_cfg": { "lr": 0.001 }, "mean": 8.592086487624329e-05, "std": 7.508073923840228e-05, "per_seed": [ 0.00011502060806378722, 6.843992014182732e-05, 0.00027265894459560513, 4.1655894165160134e-05, 7.501575601054356e-05, 4.787765283253975e-05, 3.653249950730242e-05, 3.0165643693180755e-05 ], "n": 8 }, "comparison": { "delta_mean": 4.6202021394492476e-05, "idea_wins": 1, "n_pairs": 8, "per_seed_diffs": [ 6.877319174236618e-05, 1.4300712791737169e-05, 0.00024409760771959554, 3.439410647843033e-06, 3.3412270568078384e-05, 1.1311563866911456e-05, 1.2643577065318823e-06, -6.982943887123838e-06 ], "p_value": 0.03865, "mde": 6.977709898395673e-05, "mde_rel_pct": 175.67756980642312, "verdict": "idea worse (significant)", "system_worked": false }, "math_check": { "joint_min_eig": 0.542706194000172, "schur_min_eig": 0.6700584265307041, "psd": true }, "idea_sweep": [ { "cfg": { "lr": 0.001 }, "mean": 8.592086487624329e-05 }, { "cfg": { "lr": 0.003 }, "mean": 0.00017516915022497415 }, { "cfg": { "lr": 0.006 }, "mean": 0.0001611893669632991 } ], "mechanism_signature": { "prediction": "joint Gaussian head yields calibrated predictive variance; 95% interval coverage near nominal 0.95", "observed_nll": -2.5754659175872803, "observed_coverage95": 1.0, "observed_mean_variance": 0.0008512148051522672, "confirmed": true } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_runner.py", "files": [ "bench_runner.py", "bench_report.json" ], "limitations": "The benchmark used the built-in one-step actuated-pendulum dynamics task rather than a multi-step trajectory-output task, because that is the available matched dynamics track. The Gaussian head predicted scalar next-state uncertainty, not a full future trajectory covariance; uncertainty metrics were measured on seed 0 only and were not the primary comparison metric. No low-rank covariance scaling or distribution-shift evaluation was performed.", "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }