{ "bench_version": 1, "track": "geometric_triangle_area", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.003, "budget": 8, "router": "random" }, "sweep": [ { "cfg": { "lr": 0.001, "budget": 8, "router": "random" }, "mean": 0.8043327629566193 }, { "cfg": { "lr": 0.003, "budget": 8, "router": "random" }, "mean": 0.6060835421085358 }, { "cfg": { "lr": 0.01, "budget": 8, "router": "random" }, "mean": 0.623095691204071 } ], "full": { "mean": 0.5995706543326378, "std": 0.11964870172066801, "per_seed": [ 0.6045869588851929, 0.6675567626953125, 0.846964955329895, 0.4768231213092804, 0.6347360014915466, 0.45020803809165955, 0.5004843473434448, 0.6152050495147705 ], "n": 8 } }, "idea": { "mean": 0.49004776030778885, "std": 0.12742773364281546, "per_seed": [ 0.6025350689888, 0.37475037574768066, 0.6406871676445007, 0.4309156835079193, 0.694169282913208, 0.3271963596343994, 0.40124213695526123, 0.4488860070705414 ], "n": 8 }, "comparison": { "delta_mean": -0.10952289402484894, "idea_wins": 7, "n_pairs": 8, "per_seed_diffs": [ -0.0020518898963928223, -0.29280638694763184, -0.2062777876853943, -0.045907437801361084, 0.05943328142166138, -0.12301167845726013, -0.0992422103881836, -0.16631904244422913 ], "p_value": 0.0388, "mde": 0.09513217102878818, "mde_rel_pct": 15.866715680852767, "verdict": "idea better (significant)", "system_worked": true }, "custom_track": { "name": "geometric_triangle_area", "file": "geometric_track.py", "domain": "geometric_graph" }, "idea_sweep": [ { "cfg": { "lr": 0.001, "budget": 8, "router": "greedy_jacobian" }, "mean": 0.7678551822900772 }, { "cfg": { "lr": 0.003, "budget": 8, "router": "greedy_jacobian" }, "mean": 0.49524444714188576 }, { "cfg": { "lr": 0.01, "budget": 8, "router": "greedy_jacobian" }, "mean": 0.49004776030778885 } ], "mechanism_signature": { "trained_seed": 0, "baseline_selected_gradient": 0.20941458642482758, "baseline_omitted_gradient": 0.1311078518629074, "idea_selected_gradient": 0.17618075013160706, "idea_omitted_gradient": 0.12845860421657562, "predicted_gain": "selected simplex coordinates should carry larger trained output sensitivity than omitted coordinates", "confirmed": true }, "protocol_notes": { "epochs": 18, "batch": 128, "budget": 8, "candidates": 32, "same_architecture": true, "full_8_paired_seeds": true } }