{ "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "gamma": 0.2 }, "sweep": [ { "cfg": { "lr": 0.001, "gamma": 0.2 }, "mean": 0.004167803970631212 }, { "cfg": { "lr": 0.001, "gamma": 0.5 }, "mean": 0.004167803970631212 }, { "cfg": { "lr": 0.001, "gamma": 0.8 }, "mean": 0.004167803970631212 }, { "cfg": { "lr": 0.003, "gamma": 0.2 }, "mean": 0.0035580022376962006 }, { "cfg": { "lr": 0.003, "gamma": 0.5 }, "mean": 0.0035580022376962006 }, { "cfg": { "lr": 0.003, "gamma": 0.8 }, "mean": 0.0035580022376962006 }, { "cfg": { "lr": 0.01, "gamma": 0.2 }, "mean": 0.0017142038705060259 }, { "cfg": { "lr": 0.01, "gamma": 0.5 }, "mean": 0.0017142038705060259 }, { "cfg": { "lr": 0.01, "gamma": 0.8 }, "mean": 0.0017142038705060259 } ], "full": { "mean": 0.0015303929612855427, "std": 0.0006243748661918005, "per_seed": [ 0.0019577958155423403, 0.0012857802212238312, 0.0008496360969729722, 0.00276360334828496, 0.0014558577677235007, 0.0009643874363973737, 0.0009531342657282948, 0.002012948738411069 ], "n": 8 } }, "idea": { "mean": 4.955958911523339e-05, "std": 2.2175200246237042e-05, "per_seed": [ 2.1020921849412844e-05, 4.073749732924625e-05, 2.979358032462187e-05, 4.908094342681579e-05, 5.829468136653304e-05, 4.280764551367611e-05, 9.952809341484681e-05, 5.521334969671443e-05 ], "n": 8 }, "comparison": { "delta_mean": -0.0014808333721703093, "idea_wins": 8, "n_pairs": 8, "per_seed_diffs": [ -0.0019367748936929274, -0.001245042723894585, -0.0008198425166483503, -0.002714522404858144, -0.0013975630863569677, -0.0009215797908836976, -0.000853606172313448, -0.0019577353887143545 ], "p_value": 0.0081, "mde": 0.0005620586151503473, "mde_rel_pct": 36.72642447846947, "verdict": "idea better (significant)", "system_worked": true }, "mechanism_signature": { "mechanism_signature": { "n_models": 8, "mean_abs_residual_norm": 0.7707589045166969, "mean_incremental_residual_norm": 0.0819682078436017, "mean_abs_inc_disagreement": 0.8208163678646088, "predicted_relation": "incremental residual smaller than absolute residual under persistent mismatch", "confirmed": true, "note": "NN-scale behavior measured on trained dynamics models; built-in track has no explicit injected actuator-bias condition." }, "idea_sweep": [ { "cfg": { "lr": 0.001, "gamma": 0.2 }, "mean": 0.002574956219177693 }, { "cfg": { "lr": 0.001, "gamma": 0.5 }, "mean": 0.0007634652483829996 }, { "cfg": { "lr": 0.001, "gamma": 0.8 }, "mean": 9.187122088860633e-05 }, { "cfg": { "lr": 0.003, "gamma": 0.2 }, "mean": 0.001222496688569663 }, { "cfg": { "lr": 0.003, "gamma": 0.5 }, "mean": 0.00027883463826583466 }, { "cfg": { "lr": 0.003, "gamma": 0.8 }, "mean": 8.01967735242215e-05 }, { "cfg": { "lr": 0.01, "gamma": 0.2 }, "mean": 0.0005614861802314408 }, { "cfg": { "lr": 0.01, "gamma": 0.5 }, "mean": 0.0001608945794941974 }, { "cfg": { "lr": 0.01, "gamma": 0.8 }, "mean": 4.955958911523339e-05 } ], "structural_match": "actuated pendulum rollout/control and stability" } }