{ "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006, "lambda": 0.0 }, "sweep": [ { "cfg": { "lr": 0.0015, "lambda": 0.0 }, "mean": 0.5065934360027313 }, { "cfg": { "lr": 0.003, "lambda": 0.0 }, "mean": 0.264720531180501 }, { "cfg": { "lr": 0.006, "lambda": 0.0 }, "mean": 0.05216603260487318 } ], "full": { "mean": 0.04864278854802251, "std": 0.028737614707129207, "per_seed": [ 0.02175583690404892, 0.0447821244597435, 0.11459071189165115, 0.02753545716404915, 0.05169341713190079, 0.06664472073316574, 0.03892377018928528, 0.02321626991033554 ], "n": 8 } }, "idea": { "mean": 0.051210810197517276, "std": 0.027750063002216954, "per_seed": [ 0.02233128994703293, 0.04393337294459343, 0.1131841167807579, 0.02899443544447422, 0.06591296941041946, 0.06486307084560394, 0.038853198289871216, 0.0316140279173851 ], "n": 8 }, "comparison": { "delta_mean": 0.002568021649494767, "idea_wins": 4, "n_pairs": 8, "per_seed_diffs": [ 0.0005754530429840088, -0.0008487515151500702, -0.0014065951108932495, 0.0014589782804250717, 0.014219552278518677, -0.001781649887561798, -7.05718994140625e-05, 0.00839775800704956 ], "p_value": 0.3862, "mde": 0.00477516318712563, "mde_rel_pct": 9.816795725868714, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "quantity": "median test-set G_H=max hidden separation / initial perturbation norm", "H": 8, "predicted": "rollout penalty reduces finite-horizon gain", "baseline_mean_GH": 0.5347721576690674, "idea_mean_GH": 0.5261865258216858, "observed_relative_reduction": 0.01605474728677747, "predicted_vs_observed": { "predicted_direction": "decrease", "observed_direction": "decrease" }, "confirmed": false, "paired_seed_GH_baseline": [ 0.5360813736915588, 0.5294458866119385, 0.509493350982666, 0.5483631491661072, 0.5446451306343079, 0.5424107909202576, 0.5554588437080383, 0.5122787356376648 ], "paired_seed_GH_idea": [ 0.5199079513549805, 0.5318974256515503, 0.4956446886062622, 0.5407124161720276, 0.5468658208847046, 0.5303430557250977, 0.54232257604599, 0.5017982721328735 ], "idea_config": { "lr": 0.006, "lambda": 0.01 }, "track_choice_justification": "dynamics is the built-in control/stability track with multi-step pendulum rollouts." }, "idea_sweep": [ { "cfg": { "lr": 0.006, "lambda": 0.01 }, "mean": 0.051210810197517276, "per_seed": [ 0.02233128994703293, 0.04393337294459343, 0.1131841167807579, 0.02899443544447422, 0.06591296941041946, 0.06486307084560394, 0.038853198289871216, 0.0316140279173851 ] }, { "cfg": { "lr": 0.006, "lambda": 0.03 }, "mean": 0.051318740006536245, "per_seed": [ 0.02230515517294407, 0.04414716735482216, 0.11361950635910034, 0.028880951926112175, 0.06606191396713257, 0.06464919447898865, 0.03913993015885353, 0.03174610063433647 ] }, { "cfg": { "lr": 0.006, "lambda": 0.1 }, "mean": 0.05190074094571173, "per_seed": [ 0.02226673997938633, 0.0442110151052475, 0.11737498641014099, 0.02898707240819931, 0.06658705323934555, 0.06442747265100479, 0.03898325189948082, 0.032368335872888565 ] } ], "protocol_notes": { "epochs": 5, "n_train": 400, "n_test": 200, "baseline_lr_union": [ 0.0015, 0.003, 0.006 ], "idea_lambdas": [ 0.01, 0.03, 0.1 ], "cache_used": true } }