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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Implemented an end-to-end conditional RealNVP-style controller head on the structurally matched controlled-pendulum dynamics track, sharing the GRU encoder with the baseline and using the canonical training path. The exact inverse and Jacobian log-determinant remained numerically correct after training, with reconstruction error 1.19e-7 and Jacobian-logdet error 1.19e-7. However, the idea had higher test MSE than the baseline (0.000514 vs 0.000206), with a significant paired disadvantage (delta +0.000308, p=0.0163), so it did not work on this benchmark.",
  "metrics": {
    "baseline": "Dynamics rollout MSE mean 0.000205864, std 0.000059125; best lr=0.01.",
    "idea": "4-layer conditional exact-Jacobian coupling controller MSE mean 0.000513736, std 0.000215208; best tested lr=0.01.",
    "delta_mean": 0.0003078724603255978,
    "p_value": 0.0163,
    "idea_wins": 1,
    "n_pairs": 8,
    "mechanism_signature": {
      "reconstruction_max_abs": 1.1920928955078125e-07,
      "observed_mean_abs_logdet": 0.786198616027832,
      "observed_jacobian_logdet_abs_error": 1.1920928955078125e-07,
      "jacobian_sign": 1.0,
      "confirmed": true
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_run.py",
  "files": [
    "bench_run.py",
    "bench_report.json"
  ],
  "limitations": "Only the built-in dynamics track was tested, using 800 training samples, 300 test samples, 12 epochs, and a 3-point learning-rate sweep. Longer horizons, larger datasets, alternative coupling widths/depths, probabilistic NLL, calibration, and additional controller rollout metrics were not tested.",
  "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": 0.010650048730894923
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.0021922332525718957
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 0.0001847558050940279
        }
      ],
      "full": {
        "mean": 0.00020586385835485999,
        "std": 5.912498369648286e-05,
        "per_seed": [
          0.00017305620713159442,
          0.00018997277948074043,
          0.00013568498252425343,
          0.0002403092512395233,
          0.00026725101633928716,
          0.00018888202612288296,
          0.0003143867361359298,
          0.00013736786786466837
        ],
        "n": 8
      },
      "idea_union_sweep": [
        {
          "cfg": {
            "lr": 0.01,
            "K": 4
          },
          "mean": 0.000600305080297403
        },
        {
          "cfg": {
            "lr": 0.001,
            "K": 4
          },
          "mean": 0.0048645936767570674
        },
        {
          "cfg": {
            "lr": 0.01,
            "K": 4
          },
          "mean": 0.000600305080297403
        }
      ]
    },
    "idea": {
      "mean": 0.0005137363186804578,
      "std": 0.00021520799203052408,
      "per_seed": [
        0.0005184079636819661,
        0.00034054333809763193,
        0.000662883510813117,
        0.0008793855085968971,
        0.00037677702493965626,
        0.00013013248099014163,
        0.0005593107198365033,
        0.0006424500024877489
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 0.0003078724603255978,
      "idea_wins": 1,
      "n_pairs": 8,
      "per_seed_diffs": [
        0.00034535175655037165,
        0.0001505705586168915,
        0.0005271985282888636,
        0.0006390762573573738,
        0.0001095260086003691,
        -5.874954513274133e-05,
        0.00024492398370057344,
        0.0005050821346230805
      ],
      "p_value": 0.0163,
      "mde": 0.00020001775934380757,
      "mde_rel_pct": 97.16021109398662,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "prediction": "analytic triangular inverse and logdet remain exact after training",
      "reconstruction_max_abs": 1.1920928955078125e-07,
      "observed_mean_abs_logdet": 0.786198616027832,
      "observed_jacobian_logdet_abs_error": 1.1920928955078125e-07,
      "jacobian_sign": 1.0,
      "confirmed": true
    },
    "protocol_notes": {
      "epochs": 12,
      "n_train": 800,
      "n_test": 300,
      "idea_configs": [
        {
          "lr": 0.01,
          "K": 4
        },
        {
          "lr": 0.001,
          "K": 4
        },
        {
          "lr": 0.01,
          "K": 4
        }
      ],
      "structural_match": "controlled pendulum multi-step rollout; flow is trained end-to-end with shared GRU encoder"
    }
  },
  "system_verdict": "partial",
  "practical_verdict": "harms",
  "mechanism_ok": 1,
  "system_judged": true
}