{ "bench_version": 1, "track": "dynamics", "model": "rnn_small", "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": 0.0016094102757051587 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 0.0009143295173998922 }, { "cfg": { "lr": 0.01, "weight_decay": 0.0 }, "mean": 0.0004333868491812609 } ], "full": { "mean": 0.00040872101453715004, "std": 7.309234212470366e-05, "per_seed": [ 0.0003514306154102087, 0.00042900562402792275, 0.0004498512134887278, 0.0005032599437981844, 0.00025691225891932845, 0.0003992249839939177, 0.00039930528146214783, 0.0004807781951967627 ], "n": 8 }, "grid_union": [ { "lr": 0.001, "weight_decay": 0.0 }, { "lr": 0.003, "weight_decay": 0.0 }, { "lr": 0.01, "weight_decay": 0.0 } ] }, "idea": { "mean": 0.000397826104745036, "std": 0.00010231631988854813, "per_seed": [ 0.0002917043457273394, 0.0005510751507245004, 0.00044953575707040727, 0.00048635262646712363, 0.00025691225891932845, 0.0002804143587127328, 0.00039916730020195246, 0.0004674470401369035 ], "n": 8, "sweep_grid": [ { "lr": 0.001, "weight_decay": 0.0, "kappa": 0.5 }, { "lr": 0.003, "weight_decay": 0.0, "kappa": 0.5 }, { "lr": 0.01, "weight_decay": 0.0, "kappa": 0.5 } ] }, "comparison": { "delta_mean": -1.089490979211405e-05, "idea_wins": 6, "n_pairs": 8, "per_seed_diffs": [ -5.9726269682869315e-05, 0.00012206952669657767, -3.154564183205366e-07, -1.6907317331060767e-05, 0.0, -0.00011881062528118491, -1.379812601953745e-07, -1.3331155059859157e-05 ], "p_value": 0.58435, "mde": 5.6621048472105967e-05, "mde_rel_pct": 13.853226640725977, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "track_match": "stability/control -> actuated pendulum dynamics", "idea_cfg": { "lr": 0.01, "weight_decay": 0.0, "kappa": 0.5 }, "mechanism_signature": { "residual_norm": 0.09767141938209534, "sqrt_q": 14.045147895812988, "delta": 0.006954104173928499, "gain": 1.0, "kappa": 0.5, "prediction": "delta>kappa implies a=kappa/(delta+eps) and normalized residual forcing is capped", "observed_throttled_fraction": 0.023809523809523808, "confirmed": false } } }