Polyconvex rotation-frame Jacobian loss / report_bench_2026-09-03T012737.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "The registered deformation_registration benchmark confirmed the predicted stability mechanism: the polyconvex rotation-frame loss reduced mean Jacobian condition number from 9.6255 to 3.6457, reduced fold fraction from 5.5% to 0%, and improved mean minimum determinant from -0.0314 to 0.1291. However, test MSE worsened by 0.00736 with permutation p=0.0321, yielding the required verdict idea worse (significant), so this is not a benchmark win.", "metrics": { "baseline_mean_test_mse": 0.136754, "idea_mean_test_mse": 0.144109, "delta_mean": 0.007355469278991222, "permutation_p_value": 0.0321, "math_check": { "cofactor_max_abs": 7.152557373046875e-07, "det_max_abs": 4.76837158203125e-07, "rotation_det_max_abs": 2.384185791015625e-07 }, "mechanism_signature": { "confirmed": true, "baseline_mean_condition": 9.625483721494675, "idea_mean_condition": 3.6457439363002777, "baseline_mean_min_det": -0.031383185618324205, "idea_mean_min_det": 0.1291193552315235, "baseline_fold_fraction": 0.05500000016763806, "idea_fold_fraction": 0.0 } }, "bench_report": { "bench_version": 1, "track": "deformation_registration", "model": "coordinate_warp", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_config": { "lr": 0.01, "epochs": 8 }, "sweep": { "0.001": [ 0.27781713008880615, 0.3285594880580902, 0.29619064927101135, 0.34113916754722595, 0.34860163927078247, 0.3542182743549347, 0.31516358256340027, 0.28730571269989014 ], "0.003": [ 0.2060280293226242, 0.23130172491073608, 0.2125881314277649, 0.24440883100032806, 0.2516593039035797, 0.2537933588027954, 0.2329125702381134, 0.2088773399591446 ], "0.01": [ 0.1190430298447609, 0.13602468371391296, 0.14200977981090546, 0.1490132212638855, 0.14526495337486267, 0.18109995126724243, 0.1357964277267456, 0.14082112908363342 ] }, "full": { "per_seed": [ 0.1190430298447609, 0.13602468371391296, 0.14200977981090546, 0.1490132212638855, 0.14526495337486267, 0.18109995126724243, 0.1357964277267456, 0.14082112908363342 ] } }, "idea": { "best_config": { "lr": 0.01, "epochs": 8 }, "sweep": { "0.001": [ 0.27786731719970703, 0.3285939395427704, 0.29622170329093933, 0.34119054675102234, 0.3486487567424774, 0.35429301857948303, 0.31522244215011597, 0.2873333990573883 ], "0.003": [ 0.2074858844280243, 0.2321665734052658, 0.21403560042381287, 0.24621228873729706, 0.2530094385147095, 0.25567081570625305, 0.23396754264831543, 0.21000193059444427 ], "0.01": [ 0.1227923184633255, 0.14947034418582916, 0.13802805542945862, 0.15367528796195984, 0.1728336662054062, 0.18583709001541138, 0.14342737197875977, 0.14185279607772827 ] }, "per_seed": [ 0.1227923184633255, 0.14947034418582916, 0.13802805542945862, 0.15367528796195984, 0.1728336662054062, 0.18583709001541138, 0.14342737197875977, 0.14185279607772827 ] }, "comparison": { "delta_mean": 0.007355469278991222, "idea_wins": 1, "n_pairs": 8, "per_seed_diffs": [ 0.0037492886185646057, 0.013445660471916199, -0.003981724381446838, 0.004662066698074341, 0.027568712830543518, 0.004737138748168945, 0.00763094425201416, 0.0010316669940948486 ], "p_value": 0.0321, "mde": 0.008003765508655707, "mde_rel_pct": 5.572327803121244, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "prediction": "lifted U/Cof(U)/det penalty improves conditioning and minimum determinant under large deformation", "observed_baseline_mean_condition": 9.625483721494675, "observed_idea_mean_condition": 3.6457439363002777, "observed_baseline_min_det": -0.031383185618324205, "observed_idea_min_det": 0.1291193552315235, "observed_baseline_fold_fraction": 0.05500000016763806, "observed_idea_fold_fraction": 0.0, "confirmed": true }, "custom_track": { "name": "deformation_registration", "file": "custom_deformation_track.py", "domain": "geometry" } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 run_bench.py", "files": [ "custom_deformation_track.py", "run_bench.py", "bench_results.json", "bench_stdout_registered.txt" ], "limitations": "The registered track is a synthetic deformation-registration task rather than a real clinical or graphics dataset. Rotation curl smoothing, learned input-convex energy parameters, larger models, and longer training were not tested. CUDA encountered a native illegal-instruction crash, so the completed protocol ran on CPU.", "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }