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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built and evaluated the horizon-dependent error-tube regularized GRU on the registered built-in dynamics track. The idea improved mean test MSE from 0.00052839 to 0.00046642, but the paired permutation test was not significant (p=0.19985), so the required benchmark verdict is no significant win. The trained-model mechanism signature was quantitatively confirmed, with observed/predicted hidden-change ratio 1.582.",
  "metrics": {
    "baseline": "8-seed mean test MSE 0.0005283926293; tuned lr=0.006, weight_decay=0",
    "idea": "8-seed mean test MSE 0.0004664150802; matched GRU tube penalty, best shared lr=0.006, weight_decay=0; delta=-0.0000619775492, p=0.19985"
  },
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "rnn_small",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.006,
        "weight_decay": 0.0
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "weight_decay": 0.0
          },
          "mean": 0.0013744574098382145
        },
        {
          "cfg": {
            "lr": 0.003,
            "weight_decay": 0.0
          },
          "mean": 0.000780427159043029
        },
        {
          "cfg": {
            "lr": 0.006,
            "weight_decay": 0.0
          },
          "mean": 0.0005478621314978227
        },
        {
          "cfg": {
            "lr": 0.003,
            "weight_decay": 0.0001
          },
          "mean": 0.000764783937484026
        }
      ],
      "full": {
        "mean": 0.0005283926293486729,
        "std": 7.134857758396706e-05,
        "per_seed": [
          0.0005298908217810094,
          0.000657333992421627,
          0.0004259757697582245,
          0.0005782479420304298,
          0.0004904285888187587,
          0.0005300926859490573,
          0.0004412077250890434,
          0.0005739635089412332
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.00046641508015454747,
      "std": 0.00013948062290217747,
      "per_seed": [
        0.0004731847729999572,
        0.000537174753844738,
        0.0004555774212349206,
        0.0007837752345949411,
        0.0003692372119985521,
        0.0003401216818019748,
        0.00031041333568282425,
        0.00046183622907847166
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": -6.197754919412546e-05,
      "idea_wins": 6,
      "n_pairs": 8,
      "per_seed_diffs": [
        -5.670604878105223e-05,
        -0.00012015923857688904,
        2.9601651476696134e-05,
        0.0002055272925645113,
        -0.00012119137682020664,
        -0.00018997100414708257,
        -0.00013079438940621912,
        -0.0001121272798627615
      ],
      "p_value": 0.19985,
      "mde": 0.00010505434796975666,
      "mde_rel_pct": 19.88187233028831,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "observed_hidden_change_mean": 0.03364631673321128,
      "predicted_tube_mean": 0.021336639765650034,
      "ratio_observed_over_predicted": 1.5819950263955864,
      "confirmed": true
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.006,
          "weight_decay": 0.0
        },
        "mean": 0.00046641508015454747
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 0.0007662241696380079
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 0.0014701381151098758
      }
    ],
    "idea_best_cfg": {
      "lr": 0.006,
      "weight_decay": 0.0
    },
    "protocol_notes": "Baseline and idea use matched 64-unit GRU systems, same data, epochs, batch, optimizer, and shared lr/weight-decay grid; idea tested at three shared settings."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_tube_bench.py",
  "files": [
    "stage2_tube_bench.py",
    "bench_report.json",
    "rerun_output.txt"
  ],
  "limitations": "Only the registered built-in dynamics track was tested. The GRU Jacobian estimate is a practical row-sum proxy rather than an exact autodiff/interval bound; the MVP did not test longer-horizon forecasting, explicit output constraint tightening, or violation-rate reduction. The auxiliary signature probe used CPU after a shared-GPU cuDNN allocation failure, while benchmark training retained the harness fallback.",
  "system_verdict": "partial",
  "practical_verdict": "inconclusive",
  "mechanism_ok": 1,
  "system_judged": true
}