{ "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "lam": 0.01, "epochs": 10 }, "sweep": [ { "cfg": { "lr": 0.001, "lam": 0.001, "epochs": 10 }, "mean": 0.0017361862410325557 }, { "cfg": { "lr": 0.001, "lam": 0.01, "epochs": 10 }, "mean": 0.0017254647973459214 }, { "cfg": { "lr": 0.003, "lam": 0.001, "epochs": 10 }, "mean": 0.0011354178132023662 }, { "cfg": { "lr": 0.003, "lam": 0.01, "epochs": 10 }, "mean": 0.0011304349318379536 }, { "cfg": { "lr": 0.01, "lam": 0.001, "epochs": 10 }, "mean": 0.0006082321124267764 }, { "cfg": { "lr": 0.01, "lam": 0.01, "epochs": 10 }, "mean": 0.0006044723413651809 } ], "full": { "mean": 0.0005673739506164566, "std": 0.00016282064525448582, "per_seed": [ 0.00044074832112528384, 0.0007072906009852886, 0.0003835938696283847, 0.0008862565737217665, 0.00040746666491031647, 0.00047172914491966367, 0.0006375783705152571, 0.0006043280591256917 ], "n": 8 }, "method": "pointwise constant selected-coordinate log-Jacobian" }, "idea": { "mean": 0.0003382966906428919, "std": 0.0002807304207947017, "per_seed": [ 0.00012169728142907843, 0.0010539937065914273, 0.00020155370293650776, 0.00022586695558857173, 0.00014386164548341185, 0.0003415524843148887, 0.000293777440674603, 0.00032407030812464654 ], "n": 8, "method": "cohomological potential residual" }, "comparison": { "delta_mean": -0.00022907725997356465, "idea_wins": 7, "n_pairs": 8, "per_seed_diffs": [ -0.0003190510396962054, 0.0003467031056061387, -0.00018204016669187695, -0.0006603896181331947, -0.0002636050194269046, -0.00013017666060477495, -0.00034380092984065413, -0.0002802577510010451 ], "p_value": 0.07245, "mde": 0.0002353602814500899, "mde_rel_pct": 41.482391145093814, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "mechanism_signature": { "prediction": "finite-horizon residual per step remains bounded rather than accumulating linearly", "observed": [ { "k": 1, "observed_std_R_over_k": 0.05008017644286156, "observed_std_ell_minus_c": 0.051468078047037125 }, { "k": 4, "observed_std_R_over_k": 0.05144515261054039, "observed_std_ell_minus_c": 0.052239760756492615 }, { "k": 8, "observed_std_R_over_k": 0.05219043791294098, "observed_std_ell_minus_c": 0.05265187472105026 } ], "confirmed": true }, "idea_config_sweep": [ { "cfg": { "lr": 0.01, "lam": 0.01, "epochs": 10 }, "result": { "mean": 0.0003382966906428919, "std": 0.0002807304207947017, "per_seed": [ 0.00012169728142907843, 0.0010539937065914273, 0.00020155370293650776, 0.00022586695558857173, 0.00014386164548341185, 0.0003415524843148887, 0.000293777440674603, 0.00032407030812464654 ], "n": 8, "method": "cohomological potential residual" } }, { "cfg": { "lr": 0.001, "lam": 0.01, "epochs": 10 }, "result": { "mean": 0.003544457780662924, "std": 0.0019988002357774577, "per_seed": [ 0.002059740712866187, 0.006557196378707886, 0.0014491339679807425, 0.0023036343045532703, 0.0019505396485328674, 0.00565938837826252, 0.0023294156417250633, 0.006046613212674856 ], "n": 8 } }, { "cfg": { "lr": 0.01, "lam": 0.01, "epochs": 10 }, "result": { "mean": 0.0003382966906428919, "std": 0.0002807304207947017, "per_seed": [ 0.00012169728142907843, 0.0010539937065914273, 0.00020155370293650776, 0.00022586695558857173, 0.00014386164548341185, 0.0003415524843148887, 0.000293777440674603, 0.00032407030812464654 ], "n": 8 } } ], "device": "cuda", "primary_metric": "test MSE" }, "protocol_notes": "8 paired seeds; baseline sweep uses 4 seeds then full reevaluation; shared lr/lambda union" }