Стенд-проверка (stage-2) · промт оператора:

(универсальный)

Ответ агента:

{ "worked": true, "confidence": 9, "verdict": "Implemented fractional boundary factoring with weighted quotient-gradient regularization on a custom fractional Dirichlet boundary-value regression track, using identical mlp_tiny systems and paired seeds. The idea lowered test MSE from 0.0070448734 to 0.0064753409, with paired delta -0.0005695326 and permutation p=0.0317, meeting the benchmark's significant-win criterion. The trained-model mechanism signature also confirmed the predicted near-boundary scaling.", "metrics": { "baseline": "8-seed test MSE mean=0.0070448734, std=0.0010013697; best lr=0.01, boundary_weight=1.0", "idea": "8-seed test MSE mean=0.0064753409, std=0.0010829861; best lr=0.01, lambda_g=0.01; delta=-0.0005695326; p=0.0317" }, "bench_report": { "bench_version": 1, "track": "poisson_boundary", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "boundary_weight": 1.0 }, "full": { "mean": 0.007044873433187604, "std": 0.00100136969136606, "per_seed": [ 0.006205223500728607, 0.00873982161283493, 0.008036286570131779, 0.00764721492305398, 0.005511410068720579, 0.006294506136327982, 0.0072539327666163445, 0.00667059188708663 ], "n": 8 }, "sweep": [ { "cfg": { "lr": 0.001, "boundary_weight": 1.0 }, "mean": 0.010570373153313994 }, { "cfg": { "lr": 0.001, "boundary_weight": 10.0 }, "mean": 0.014104659901931882 }, { "cfg": { "lr": 0.001, "boundary_weight": 30.0 }, "mean": 0.05715882801450789 }, { "cfg": { "lr": 0.003, "boundary_weight": 1.0 }, "mean": 0.008062692591920495 }, { "cfg": { "lr": 0.003, "boundary_weight": 10.0 }, "mean": 0.010542358737438917 }, { "cfg": { "lr": 0.003, "boundary_weight": 30.0 }, "mean": 0.014223965117707849 }, { "cfg": { "lr": 0.01, "boundary_weight": 1.0 }, "mean": 0.007657136651687324 }, { "cfg": { "lr": 0.01, "boundary_weight": 10.0 }, "mean": 0.01242852327413857 }, { "cfg": { "lr": 0.01, "boundary_weight": 30.0 }, "mean": 0.04939912143163383 } ] }, "idea": { "best_cfg": { "lr": 0.01, "lambda_g": 0.01 }, "per_seed": [ 0.005347711965441704, 0.007960012182593346, 0.00744650699198246, 0.006636394187808037, 0.004956180229783058, 0.005183100700378418, 0.007344032172113657, 0.006928788498044014 ], "mean": 0.006475340866018087, "std": 0.0010829860742993793, "n": 8 }, "comparison": { "delta_mean": -0.0005695325671695173, "idea_wins": 6, "n_pairs": 8, "per_seed_diffs": [ -0.0008575115352869034, -0.0007798094302415848, -0.0005897795781493187, -0.001010820735245943, -0.000555229838937521, -0.001111405435949564, 9.009940549731255e-05, 0.0002581966109573841 ], "p_value": 0.0317, "mde": 0.0004164460873821372, "mde_rel_pct": 5.911335261472642, "verdict": "idea better (significant)", "system_worked": true }, "custom_track": { "name": "poisson_boundary", "file": "fractional_dirichlet_track.py", "domain": "pde" }, "idea_sweep": [ { "cfg": { "lr": 0.001, "lambda_g": 0.0 }, "mean": 0.013011600589379668 }, { "cfg": { "lr": 0.003, "lambda_g": 0.001 }, "mean": 0.008521844982169569 }, { "cfg": { "lr": 0.01, "lambda_g": 0.01 }, "mean": 0.0068476563319563866 } ], "mechanism_signature": { "quantity": "median absolute trained prediction at d=0.02 divided by median absolute trained prediction at d=0.10", "predicted_idea_ratio": 0.2990697562442441, "observed_idea": 0.302967306501504, "observed_baseline": 0.37791017847441105, "confirmed": true } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "fractional_dirichlet_track.py", "bench_report.json" ], "limitations": "The benchmark uses a manufactured fractional-boundary regression proxy rather than an explicit fractional operator quadrature/PINN residual. Only a unit-disk domain, a=0.75, one small MLP architecture, 35 epochs, and 400 training/test samples were evaluated; no higher-dimensional, nonlocal-operator, or broader architecture study was performed.", "system_verdict": "worked", "practical_verdict": "helps", "mechanism_ok": 1, "system_judged": true }