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

(универсальный)

Ответ агента:

{ "worked": true, "confidence": 7, "verdict": "Implemented a matched dynamics-track comparison using the shared rnn_small architecture. The intervention projects the recurrent matrix after each optimizer step to a certified spectral contraction bound and significantly improved test MSE over the tuned vanilla baseline on all 8 paired seeds. The Lipschitz/contraction sanity check passed, and the trained-model perturbation sensitivity stayed below the measured recurrent-norm bound.", "metrics": { "baseline": "Tuned vanilla rnn_small via bench.train_model, lr=0.006: mean test MSE 0.0006132683593, std 0.0001242455317.", "idea": "Projected contractive rnn_small, lr=0.006, recurrent norm bound 0.88: mean test MSE 0.0001349421136, std 0.0000230953576.", "delta_mean": -0.0004783262457, "relative_improvement_pct": 77.998, "p_value": 0.0081, "idea_wins": 8, "n_pairs": 8 }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "bench_report.json" ], "limitations": "This is a conservative spectral projection proxy for the proposed quadratic Lyapunov certificate, not a general data-driven SDP with trajectory reconstruction. Only the built-in controlled-pendulum dynamics track was tested with 400 training samples, 400 test samples, 20 epochs, one radius, and a three-value learning-rate grid. The mechanism signature used output sensitivity to an input perturbation rather than direct hidden-state free-response decay or disturbance-to-output gain.", "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": 0.0014035784406587481 }, { "cfg": { "lr": 0.003 }, "mean": 0.000857145045301877 }, { "cfg": { "lr": 0.006 }, "mean": 0.0005846045824000612 } ], "full": { "mean": 0.0006132683593023103, "std": 0.00012424553172221234, "per_seed": [ 0.0005031590699218214, 0.0007117226487025619, 0.0004876864841207862, 0.0006358501268550754, 0.0004713028611149639, 0.0005687447264790535, 0.0006691908929497004, 0.0008584900642745197 ], "n": 8 } }, "idea": { "mean": 0.0001349421136183082, "std": 2.309535756281096e-05, "per_seed": [ 0.00011737410386558622, 0.00014533710782416165, 0.0001241672143805772, 0.0001715633989078924, 0.0001070009166141972, 0.00010650144395185634, 0.00014698794984724373, 0.00016060477355495095 ], "n": 8 }, "comparison": { "delta_mean": -0.00047832624568400206, "idea_wins": 8, "n_pairs": 8, "per_seed_diffs": [ -0.00038578496605623513, -0.0005663855408784002, -0.000363519269740209, -0.00046428672794718295, -0.0003643019445007667, -0.00046224328252719715, -0.0005222029431024566, -0.0006978852907195687 ], "p_value": 0.0081, "mde": 9.642354488901526e-05, "mde_rel_pct": 15.722895764378306, "verdict": "idea better (significant)", "system_worked": true }, "mechanism_signature": { "math_sanity": { "max_ratio": 0.7304509862882015, "bound": 0.88, "passed": true }, "selected_cfg": { "lr": 0.006, "radius": 0.88 }, "candidate_results": [ { "cfg": { "lr": 0.001, "radius": 0.88 }, "mean": 0.0023357835689239437 }, { "cfg": { "lr": 0.003, "radius": 0.88 }, "mean": 0.0002383494229434291 }, { "cfg": { "lr": 0.006, "radius": 0.88 }, "mean": 0.0001349421136183082 } ], "predicted_contraction_bound": 0.8710182309150696, "observed_input_sensitivity": 0.007450580131262541, "trained_recurrent_norm": 0.8710182309150696, "within_bound": true, "confirmed": true } }, "system_verdict": "worked", "practical_verdict": "helps", "mechanism_ok": 1, "system_judged": true }