{ "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006, "weight_decay": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0 }, "mean": 0.012346302974037826 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 0.004180165706202388 }, { "cfg": { "lr": 0.006, "weight_decay": 0.0 }, "mean": 0.0025389266083948314 } ], "full": { "mean": 0.0024029802589211613, "std": 0.0007328567905589416, "per_seed": [ 0.001808110624551773, 0.0038746704813092947, 0.002130982931703329, 0.002341942396014929, 0.001473060343414545, 0.001914751948788762, 0.003167124465107918, 0.0025131988804787397 ], "n": 8 } }, "idea": { "mean": 0.005464809873956256, "std": 0.002061025522134347, "per_seed": [ 0.007386896759271622, 0.007307976484298706, 0.006752352695912123, 0.0009749423479661345, 0.0034697847440838814, 0.006274467799812555, 0.006125933025032282, 0.005426125135272741 ], "n": 8 }, "comparison": { "delta_mean": 0.0030618296150350943, "idea_wins": 1, "n_pairs": 8, "per_seed_diffs": [ 0.005578786134719849, 0.0034333060029894114, 0.004621369764208794, -0.0013670000480487943, 0.0019967244006693363, 0.004359715851023793, 0.002958808559924364, 0.0029129262547940016 ], "p_value": 0.0152, "mde": 0.0017701285805840077, "mde_rel_pct": 73.66388358840376, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "prediction": "anti-oscillation penalty should reduce trained recurrent two-step similarity relative to adjacent change", "baseline_mean_two_to_one_ratio": 2.1946372059836357, "idea_mean_two_to_one_ratio": 2.845283087236918, "predicted_direction": "idea lower than baseline", "confirmed": false, "per_seed": [ { "seed": 0, "baseline": { "observed_two_step_mse": 0.004077367950230837, "observed_adjacent_mse": 0.001899817492812872, "observed_two_to_one_ratio": 2.1461779901889475 }, "idea": { "observed_two_step_mse": 4.882162102148868e-05, "observed_adjacent_mse": 1.8978958905790932e-05, "observed_two_to_one_ratio": 2.5710530663479334 } }, { "seed": 1, "baseline": { "observed_two_step_mse": 0.0035950804594904184, "observed_adjacent_mse": 0.0015977286966517568, "observed_two_to_one_ratio": 2.250105393969814 }, "idea": { "observed_two_step_mse": 3.846972686005756e-05, "observed_adjacent_mse": 1.3402634067460895e-05, "observed_two_to_one_ratio": 2.868170910096271 } }, { "seed": 2, "baseline": { "observed_two_step_mse": 0.0043192701414227486, "observed_adjacent_mse": 0.0019986697006970644, "observed_two_to_one_ratio": 2.161061694836021 }, "idea": { "observed_two_step_mse": 3.958393062930554e-05, "observed_adjacent_mse": 1.3548745300795417e-05, "observed_two_to_one_ratio": 2.9194390595259114 } }, { "seed": 3, "baseline": { "observed_two_step_mse": 0.003866632701829076, "observed_adjacent_mse": 0.0016693203942850232, "observed_two_to_one_ratio": 2.316277661430325 }, "idea": { "observed_two_step_mse": 3.4838722058339044e-05, "observed_adjacent_mse": 1.1200490916962735e-05, "observed_two_to_one_ratio": 3.1076892454035296 } }, { "seed": 4, "baseline": { "observed_two_step_mse": 0.0038868181873112917, "observed_adjacent_mse": 0.0016667143208906054, "observed_two_to_one_ratio": 2.332010242242335 }, "idea": { "observed_two_step_mse": 3.9646634832024574e-05, "observed_adjacent_mse": 1.4838369679637253e-05, "observed_two_to_one_ratio": 2.6701001987036426 } }, { "seed": 5, "baseline": { "observed_two_step_mse": 0.0034053579438477755, "observed_adjacent_mse": 0.001578447176143527, "observed_two_to_one_ratio": 2.157396472527507 }, "idea": { "observed_two_step_mse": 4.4292191887507215e-05, "observed_adjacent_mse": 1.595341927895788e-05, "observed_two_to_one_ratio": 2.774605559968646 } }, { "seed": 6, "baseline": { "observed_two_step_mse": 0.0034227895084768534, "observed_adjacent_mse": 0.0015517541905865073, "observed_two_to_one_ratio": 2.205740749297205 }, "idea": { "observed_two_step_mse": 3.656304397736676e-05, "observed_adjacent_mse": 1.259200507774949e-05, "observed_two_to_one_ratio": 2.9013671833797035 } }, { "seed": 7, "baseline": { "observed_two_step_mse": 0.003476944053545594, "observed_adjacent_mse": 0.0017486677970737219, "observed_two_to_one_ratio": 1.9883274433769293 }, "idea": { "observed_two_step_mse": 3.67501015716698e-05, "observed_adjacent_mse": 1.2448339475668035e-05, "observed_two_to_one_ratio": 2.949839474469707 } } ] }, "idea_trials": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0, "lambda_cycle": 0.01 }, "result": { "mean": 0.008634913916466758, "std": 0.007319620916715749, "per_seed": [ 0.0085679329931736, 0.024327190592885017, 0.0032044434919953346, 0.00580877298489213, 0.003978467080742121, 0.004719811026006937, 0.016555899754166603, 0.0019167934078723192 ], "n": 8 } }, { "cfg": { "lr": 0.003, "weight_decay": 0.0, "lambda_cycle": 0.03 }, "result": { "mean": 0.01017151348059997, "std": 0.0028825955800251195, "per_seed": [ 0.009243861772119999, 0.009108055382966995, 0.013492719270288944, 0.015129592269659042, 0.010461220517754555, 0.009117119945585728, 0.004917753394693136, 0.009901785291731358 ], "n": 8 } }, { "cfg": { "lr": 0.006, "weight_decay": 0.0, "lambda_cycle": 0.1 }, "result": { "mean": 0.005464809873956256, "std": 0.002061025522134347, "per_seed": [ 0.007386896759271622, 0.007307976484298706, 0.006752352695912123, 0.0009749423479661345, 0.0034697847440838814, 0.006274467799812555, 0.006125933025032282, 0.005426125135272741 ], "n": 8 } } ], "structural_match": "Dynamics track: recurrent GRU predicts actuated pendulum rollout; anti-oscillation targets recurrent hidden-state period-two behaviour.", "protocol": { "paired_seeds": [ 0, 1, 2, 3, 4, 5, 6, 7 ], "epochs": 12, "batch": 128, "baseline_grid": [ { "lr": 0.001, "weight_decay": 0.0 }, { "lr": 0.003, "weight_decay": 0.0 }, { "lr": 0.006, "weight_decay": 0.0 } ], "idea_grid": [ { "lr": 0.001, "weight_decay": 0.0, "lambda_cycle": 0.01 }, { "lr": 0.003, "weight_decay": 0.0, "lambda_cycle": 0.03 }, { "lr": 0.006, "weight_decay": 0.0, "lambda_cycle": 0.1 } ] } }