{ "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.0045, "temp": 1.5 }, "sweep": [ { "cfg": { "lr": 0.0015, "temp": 0.5 }, "mean": 0.0014554406952811405 }, { "cfg": { "lr": 0.0015, "temp": 1.0 }, "mean": 0.0010139688311028294 }, { "cfg": { "lr": 0.0015, "temp": 1.5 }, "mean": 0.0007530815055361018 }, { "cfg": { "lr": 0.003, "temp": 0.5 }, "mean": 0.0010139688311028294 }, { "cfg": { "lr": 0.003, "temp": 1.0 }, "mean": 0.0006243628085940145 }, { "cfg": { "lr": 0.003, "temp": 1.5 }, "mean": 0.00043806652320199646 }, { "cfg": { "lr": 0.0045, "temp": 0.5 }, "mean": 0.0007530815055361018 }, { "cfg": { "lr": 0.0045, "temp": 1.0 }, "mean": 0.00043806652320199646 }, { "cfg": { "lr": 0.0045, "temp": 1.5 }, "mean": 0.00030478146436507814 } ], "full": { "mean": 0.00030478146436507814, "std": 8.963828849728333e-05, "per_seed": [ 0.00031690497417002916, 0.00044110423186793923, 0.00018014911620412022, 0.0003148825780954212, 0.00030254252487793565, 0.00016191434406209737, 0.0003196565667167306, 0.00040109737892635167 ], "n": 8 } }, "idea": { "mean": 0.00020417736959643662, "std": 7.801088951813288e-05, "per_seed": [ 0.00024111710081342608, 0.00025753368390724063, 0.00017732763080857694, 0.00028794861282221973, 0.00021670988644473255, 7.447090320056304e-05, 9.091229730984196e-05, 0.00028739884146489203 ], "n": 8 }, "comparison": { "delta_mean": -0.00010060409476864152, "idea_wins": 8, "n_pairs": 8, "per_seed_diffs": [ -7.578787335660309e-05, -0.0001835705479606986, -2.8214853955432773e-06, -2.6933965273201466e-05, -8.58326384332031e-05, -8.744344086153433e-05, -0.00022874426940688863, -0.00011369853746145964 ], "p_value": 0.0081, "mde": 6.278775263566925e-05, "mde_rel_pct": 20.600909168301598, "verdict": "idea better (significant)", "system_worked": true }, "mechanism_signature": { "n_updates": 1008, "predicted_mean_shift": 0.0022922240415369968, "observed_mean_shift": 0.009866443170303394, "rho_mean": 22.36028769385943, "rho_max": 518.6289797241246, "confirmed": false, "baseline_sweep_union": [ { "lr": 0.0015, "temp": 0.5 }, { "lr": 0.0015, "temp": 1.0 }, { "lr": 0.0015, "temp": 1.5 }, { "lr": 0.003, "temp": 0.5 }, { "lr": 0.003, "temp": 1.0 }, { "lr": 0.003, "temp": 1.5 }, { "lr": 0.0045, "temp": 0.5 }, { "lr": 0.0045, "temp": 1.0 }, { "lr": 0.0045, "temp": 1.5 } ], "idea_sweep": [ { "cfg": { "lr": 0.0045, "delta": 0.005 }, "result": { "mean": 0.0002496864208296756, "std": 0.00010117674323671589, "per_seed": [ 0.00031152486917562783, 0.0002743446093518287, 0.00018798311066348106, 0.00016755118849687278, 0.00040196749614551663, 0.00010216210648650303, 0.0003774008946493268, 0.00017455709166824818 ], "n": 8 } }, { "cfg": { "lr": 0.0045, "delta": 0.01 }, "result": { "mean": 0.00024624322350064176, "std": 9.45256265005067e-05, "per_seed": [ 0.00026604763115756214, 0.0002716899325605482, 0.00027435951051302254, 0.00038485671393573284, 0.00032344082137569785, 4.713781163445674e-05, 0.00018148659728467464, 0.00022092676954343915 ], "n": 8 } }, { "cfg": { "lr": 0.0045, "delta": 0.02 }, "result": { "mean": 0.00020417736959643662, "std": 7.801088951813288e-05, "per_seed": [ 0.00024111710081342608, 0.00025753368390724063, 0.00017732763080857694, 0.00028794861282221973, 0.00021670988644473255, 7.447090320056304e-05, 9.091229730984196e-05, 0.00028739884146489203 ], "n": 8 } } ] } }