Differentially Passive Neural Blocks / report_bench_2026-09-02T165928.md
Failed on benchmark
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented Differentially Passive Neural Blocks as a local sampled GRU Jacobian-contraction penalty on the structurally matched dynamics track, using the benchmark's rnn_small architecture and paired-seed protocol. The idea achieved lower mean MSE than the tuned baseline (0.002211 versus 0.004185; delta -0.001974; 6/8 wins), but the permutation p-value was 0.14225, so there was no significant win and the harness verdict is "no significant win." The mechanism signature was not confirmed because its diagnostic run hit a cuDNN host-allocation error, so no robustness claim is made from that signature.", "metrics": { "baseline": "dynamics/rnn_small, tuned lr=0.01; full 8-seed test MSE mean 0.00418536, std 0.00262483", "idea": "sampled GRU contraction penalty, lr=0.01; full 8-seed test MSE mean 0.00221110, std 0.00158126; paired delta -0.00197427; p=0.14225; 6/8 wins; verdict no significant win" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "bench_report.json" ], "limitations": "The benchmark used 8 epochs and 400 train/400 test samples to fit the 40-minute budget rather than the README's larger default workload. The intervention used sampled hidden states rather than a formal all-domain certificate; the P=LLᵀ parameterization and continuous-time LMI were not tested. The trained-model mechanism signature failed under a cuDNN host-allocation error and is marked confirmed=false. The idea's lower mean MSE is therefore suggestive but not statistically validated under the required permutation test.", "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01 }, "sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.02264283667318523 }, { "cfg": { "lr": 0.003 }, "mean": 0.006244247197173536 }, { "cfg": { "lr": 0.01 }, "mean": 0.00514151350944303 } ], "full": { "mean": 0.004185360565315932, "std": 0.00262482957347662, "per_seed": [ 0.0021960907615721226, 0.008501192554831505, 0.0017922910628840327, 0.008076479658484459, 0.0030377130024135113, 0.0011907307198271155, 0.0035283213946968317, 0.005160065367817879 ], "n": 8 }, "union_grid": [ { "lr": 0.001 }, { "lr": 0.003 }, { "lr": 0.01 } ] }, "idea": { "mean": 0.002211095365055371, "std": 0.001581263928471004, "per_seed": [ 0.0014746395172551274, 0.002074356423690915, 0.000778969086240977, "0.0025305193848907948", 0.002626171102747321, 0.0014305157819762826, 0.006008145399391651, 0.0007654462242498994 ], "n": 8 }, "comparison": { "delta_mean": -0.001974265200260561, "idea_wins": 6, "n_pairs": 8, "per_seed_diffs": [ -0.0007214512443169951, -0.00642683613114059, -0.0010133219766430557, -0.005545960273593664, -0.00041154189966619015, 0.00023978506214916706, 0.0024798240046948195, -0.004394619143567979 ], "p_value": 0.14225, "mde": 0.0026083344497749737, "mde_rel_pct": 62.32042398904009, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "predicted_ratio_bound": 0.9591663046625439, "observed_mean_ratio": null, "predicted_lmi_bound": 0.0, "observed_max_lmi": null, "confirmed": false, "error": "RuntimeError('cuDNN error: CUDNN_STATUS_INTERNAL_ERROR_HOST_ALLOCATION_FAILED')" }, "selected_idea_cfg": { "lr": 0.01 }, "structural_match": "Dynamics track directly tests recurrent trajectory sensitivity and long-horizon stability." }, "system_verdict": "failed", "practical_verdict": "inconclusive", "mechanism_ok": 0, "system_judged": true }