# Стенд-проверка (stage-2) · промт оператора: (универсальный) Ответ агента: { "worked": false, "confidence": 10, "verdict": "The nullspace-coordinate MLP exactly enforced the Dirichlet constraints on trained benchmark models, confirming the mechanism signature, but it significantly worsened the independent test MSE relative to the same MLP predicting the full field. The registered PDE benchmark therefore rejects the claimed overall win.", "metrics": { "baseline": "test MSE mean 0.00022453273413702846, std 0.00002826627803497815", "idea": "test MSE mean 0.0007747690397081897, std 0.00018247493264257159", "paired_delta": "+0.0005502363055711612; permutation p=0.0081; idea wins 0/8" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_stage2.py", "files": [ "bench_stage2.py", "bench_report.json" ], "limitations": "Only the registered poisson_dirichlet PDE track and mlp_tiny model were tested. Unseen geometries, recurrent rollouts, nonhomogeneous affine constraints, frozen multi-block mechanisms, runtime/FLOP overhead, and larger neural operators were not evaluated.", "bench_report": { "bench_version": 1, "track": "poisson_dirichlet", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "weight_decay": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0 }, "mean": 0.0014014299667906016 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 0.00047322634054580703 }, { "cfg": { "lr": 0.01, "weight_decay": 0.0 }, "mean": 0.00023210687140817754 } ], "full": { "mean": 0.00022453273413702846, "std": 2.826627803497815e-05, "per_seed": [ 0.00021784205455332994, 0.00023001956287771463, 0.00027551333187147975, 0.00020505253633018583, 0.00025107755209319293, 0.0002373710594838485, 0.00019424316997174174, 0.0001851426059147343 ], "n": 8 }, "idea_union_trials": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0 }, "mean": 0.00744940631557256 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 0.002477104659192264 }, { "cfg": { "lr": 0.01, "weight_decay": 0.0 }, "mean": 0.0007909908308647573 } ] }, "idea": { "mean": 0.0007747690397081897, "std": 0.00018247493264257159, "per_seed": [ 0.0009708237485028803, 0.0005817511701025069, 0.0010273662628605962, 0.0005840221419930458, 0.0007391659892164171, 0.0005464869318529963, 0.0009673092281445861, 0.0007812268449924886 ], "n": 8 }, "comparison": { "delta_mean": 0.0005502363055711612, "idea_wins": 0, "n_pairs": 8, "per_seed_diffs": [ 0.0007529816939495504, 0.00035173160722479224, 0.0007518529309891164, 0.00037896960566286, 0.00048808843712322414, 0.00030911587236914784, 0.0007730660581728444, 0.0005960842390777543 ], "p_value": 0.0081, "mde": 0.00016245867586821579, "mde_rel_pct": 72.3541164243206, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "selection": "best of shared 3-point LR/weight-decay grid on seeds 0-3", "idea_best_cfg": { "lr": 0.01, "weight_decay": 0.0 }, "prediction": "nullspace decoder should make C u exactly zero while ambient output violates endpoints", "predicted_boundary_residual_idea": 0.0, "observed_boundary_residual_idea_max": 0.0, "observed_boundary_residual_baseline_max": 0.02415449172258377, "observed_boundary_residual_baseline_mean": 0.00435631675645709, "observed_boundary_residual_idea_mean": 0.0, "ratio_baseline_to_idea_max": 24154490880.0, "confirmed": true, "trained_models": true }, "custom_track": { "name": "poisson_dirichlet", "file": "/home/maxwelhelp/all/math2nn/bench/custom_tracks/poisson_dirichlet.py", "domain": "pde", "reason": "built-in custom PDE track has fixed-grid field outputs and homogeneous Dirichlet constraints" }, "protocol": { "epochs": 20, "batch": 128, "seeds": [ 0, 1, 2, 3, 4, 5, 6, 7 ], "grid": [ { "lr": 0.001, "weight_decay": 0.0 }, { "lr": 0.003, "weight_decay": 0.0 }, { "lr": 0.01, "weight_decay": 0.0 } ] } }, "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }