{ "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.009, "epochs": 20 }, "sweep": [ { "config": { "lr": 0.001, "epochs": 20 }, "mean": 0.0014746972592547536, "per_seed": [ 0.0012720254017040133, 0.0014296569861471653, 0.0013667949242517352, 0.0018303117249161005 ] }, { "config": { "lr": 0.003, "epochs": 20 }, "mean": 0.0009116442524828017, "per_seed": [ 0.0007490376010537148, 0.000787030381616205, 0.0009362144046463072, 0.0011742946226149797 ] }, { "config": { "lr": 0.009, "epochs": 20 }, "mean": 0.00042038498941110447, "per_seed": [ 0.0002886170113924891, 0.0003179473278578371, 0.00030860063270665705, 0.0007663749856874347 ] } ], "full": { "per_seed": [ 0.0002886170113924891, 0.0003179473278578371, 0.00030860063270665705, 0.0007663749856874347, 0.0003769927134271711, 0.00034989885170944035, 0.0004854405706282705, 0.0004627566668204963 ], "mean": 0.0004195785950287245 } }, "idea": { "config": { "lr": 0.009, "epochs": 20, "comm_lambda": 0.05 }, "per_seed": [ 1.915471875690855e-05, 5.7846078561851755e-05, 4.466433892957866e-05, 1.5033640011097305e-05, 1.8127400835510343e-05, 3.6396970244823024e-05, 2.411130117252469e-05, 2.4222023057518527e-05 ], "mean": 2.9944558946226607e-05 }, "comparison": { "delta_mean": -0.0003896340360824979, "idea_wins": 8, "n_pairs": 8, "per_seed_diffs": [ -0.0002694622926355805, -0.0002601012492959853, -0.0002639362937770784, -0.0007513413456763374, -0.00035886531259166077, -0.0003135018814646173, -0.0004613292694557458, -0.0004385346437629778 ], "p_value": 0.0081, "mde": 0.00013852579096433872, "mde_rel_pct": 33.01545708137356, "verdict": "idea better (significant)", "system_worked": true }, "track_justification": "Dynamics is structurally matched to switched latent stability and Lyapunov contraction.", "mechanism_signature": { "prediction": "commutator regularization lowers trained ordered-cycle noncommutativity and cycle gain", "observed": { "observed_commutator_mean": 0.0003028972596865158, "observed_cycle_gain_mean": 0.9300689920783043, "n_models": 8 }, "baseline_observed": { "cycle_gain_mean": NaN, "input_sensitivity_mean": 0.09274659305810928, "note": "GRU has no explicit matrix-cycle generator" }, "idea_input_sensitivity_mean": 0.07906531821936369, "confirmed": false }, "idea_sweep": [ { "config": { "lr": 0.009, "epochs": 20, "comm_lambda": 0.0 }, "mean": 2.995237787217775e-05, "per_seed": [ 1.9160790543537587e-05, 5.785106623079628e-05, 4.468455517780967e-05, 1.5057028576848097e-05, 1.8133563571609557e-05, 3.6374363844515756e-05, 2.4118378860293888e-05, 2.423927617201116e-05 ] }, { "config": { "lr": 0.009, "epochs": 20, "comm_lambda": 0.01 }, "mean": 2.995091836055508e-05, "per_seed": [ 1.9159475414198823e-05, 5.7849942095344886e-05, 4.4682146835839376e-05, 1.5051737136673182e-05, 1.813237940950785e-05, 3.637870031525381e-05, 2.4117280190694146e-05, 2.423568548692856e-05 ] }, { "config": { "lr": 0.009, "epochs": 20, "comm_lambda": 0.05 }, "mean": 2.9944558946226607e-05, "per_seed": [ 1.915471875690855e-05, 5.7846078561851755e-05, 4.466433892957866e-05, 1.5033640011097305e-05, 1.8127400835510343e-05, 3.6396970244823024e-05, 2.411130117252469e-05, 2.4222023057518527e-05 ] } ] }