{ "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": 0.14925161935389042 }, { "cfg": { "lr": 0.003 }, "mean": 0.030697030015289783 }, { "cfg": { "lr": 0.006 }, "mean": 0.012412646785378456 } ], "full": { "mean": 0.01222972312825732, "std": 0.0049134211484049405, "per_seed": [ 0.014105317182838917, 0.013159467838704586, 0.00736610684543848, 0.015019695274531841, 0.003357629058882594, 0.012679333798587322, 0.02101588435471058, 0.011134350672364235 ], "n": 8 } }, "idea": { "mean": 0.012229697429575026, "std": 0.004913482894403578, "per_seed": [ 0.014105571433901787, 0.013159754686057568, 0.007366028614342213, 0.015018501318991184, 0.003357773646712303, 0.01267978735268116, 0.021016454324126244, 0.01113370805978775 ], "n": 8 }, "comparison": { "delta_mean": -2.569868229329586e-08, "idea_wins": 3, "n_pairs": 8, "per_seed_diffs": [ 2.5425106287002563e-07, 2.868473529815674e-07, -7.82310962677002e-08, -1.1939555406570435e-06, 1.4458782970905304e-07, 4.5355409383773804e-07, 5.699694156646729e-07, -6.426125764846802e-07 ], "p_value": 0.91285, "mde": 5.034535119956499e-07, "mde_rel_pct": 0.004116638673793017, "verdict": "no measurable effect", "system_worked": false }, "mechanism_signature": { "mechanism_signature": { "baseline_min_curvature": 0.049999963492155075, "idea_min_curvature": 0.04999997094273567, "baseline_trace_curvature": 1.396693229675293, "idea_trace_curvature": 1.3966948986053467, "baseline_observed_perturbation": 7.737995474599302e-05, "idea_observed_perturbation": 7.729607750661671e-05, "predicted_inverse_margin_ratio": 0.9999998211860657, "observed_perturbation_ratio": 0.9989160340083987, "confirmed": true }, "idea_trials": [ { "cfg": { "lr": 0.006 }, "result": { "mean": 0.012229697429575026, "std": 0.004913482894403578, "per_seed": [ 0.014105571433901787, 0.013159754686057568, 0.007366028614342213, 0.015018501318991184, 0.003357773646712303, 0.01267978735268116, 0.021016454324126244, 0.01113370805978775 ], "n": 8 } }, { "cfg": { "lr": 0.003 }, "result": { "mean": 0.026620963122695684, "std": 0.005782354818965008, "per_seed": [ 0.035109490156173706, 0.027120579034090042, 0.02559320256114006, 0.03496447950601578, 0.025693824514746666, 0.02174903266131878, 0.02609817497432232, 0.016638921573758125 ], "n": 8 } }, { "cfg": { "lr": 0.006 }, "result": { "mean": 0.012229697778820992, "std": 0.0049134830277391085, "per_seed": [ 0.01410557422786951, 0.013159754686057568, 0.007366028614342213, 0.015018501318991184, 0.003357773646712303, 0.01267978735268116, 0.021016454324126244, 0.01113370805978775 ], "n": 8 } } ], "structural_match": "dynamics: recurrent actuated pendulum and stability/control task", "intervention": "soft-min of input Jacobian Gauss-Newton curvature plus fixed prior" } }