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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 10,
  "verdict": "Implemented the data-driven hidden-state ellipsoid intervention on the registered dynamics track with the shared rnn_small architecture. The tuned baseline significantly outperformed the best idea configuration: baseline MSE 0.00054517 versus idea MSE 0.24843272, paired delta +0.24788754, permutation p=0.0081. The trained-model signature was not confirmed because normalized ellipsoid violation was 100% and maximum energy was 22.335 rather than the predicted bound of 1.",
  "metrics": {
    "baseline": "lr=0.01; mean test MSE=0.0005451723; std=0.0002036184",
    "idea": "lr=0.01, lambda=0.01; mean test MSE=0.2484327164; std=0.1289024479; delta=+0.2478875441; p=0.0081; wins=0"
  },
  "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": 0.001864717691205442
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.0010901302302954718
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 0.0006186682076076977
        }
      ],
      "full": {
        "mean": 0.00054517226817552,
        "std": 0.00020361842727135583,
        "per_seed": [
          0.0004930912400595844,
          0.0004765945195686072,
          0.0004673648509196937,
          0.0010376222198829055,
          0.000511663849465549,
          0.00027233464061282575,
          0.0005571767687797546,
          0.0005455300561152399
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.24843271635472775,
      "std": 0.12890244786167743,
      "per_seed": [
        0.09243064373731613,
        0.48731154203414917,
        0.30256471037864685,
        0.0956249013543129,
        0.23182319104671478,
        0.24627986550331116,
        0.155330628156662,
        0.376096248626709
      ],
      "n": 8
    },
    "idea_config": {
      "lr": 0.01,
      "lambda": 0.01
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "lambda": 0.01
        },
        "mean": 0.74566999822855
      },
      {
        "cfg": {
          "lr": 0.001,
          "lambda": 0.1
        },
        "mean": 0.7457223460078239
      },
      {
        "cfg": {
          "lr": 0.001,
          "lambda": 1.0
        },
        "mean": 0.745727613568306
      },
      {
        "cfg": {
          "lr": 0.003,
          "lambda": 0.01
        },
        "mean": 0.6492044553160667
      },
      {
        "cfg": {
          "lr": 0.003,
          "lambda": 0.1
        },
        "mean": 0.6523455232381821
      },
      {
        "cfg": {
          "lr": 0.003,
          "lambda": 1.0
        },
        "mean": 0.6526677086949348
      },
      {
        "cfg": {
          "lr": 0.01,
          "lambda": 0.01
        },
        "mean": 0.24843271635472775
      },
      {
        "cfg": {
          "lr": 0.01,
          "lambda": 0.1
        },
        "mean": 0.37393518909811974
      },
      {
        "cfg": {
          "lr": 0.01,
          "lambda": 1.0
        },
        "mean": 0.39147959649562836
      }
    ],
    "comparison": {
      "delta_mean": 0.24788754408655223,
      "p_value": 0.0081,
      "per_seed_diffs": [
        0.09193755249725655,
        0.48683494751458056,
        0.30209734552772716,
        0.09458727913442999,
        0.23131152719724923,
        0.24600753086269833,
        0.15477345138788223,
        0.37555071857059374
      ],
      "idea_wins": 0,
      "n_pairs": 8,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "prediction": "data-fitted ellipsoid barrier plus projection keeps normalized hidden energy bounded",
      "baseline_mean_hidden_energy": 2.286437749862671,
      "idea_mean_normalized_energy": 9.777409553527832,
      "idea_max_normalized_energy": 22.335248947143555,
      "idea_violation_rate": 1.0,
      "predicted_max_energy": 1.0,
      "confirmed": false
    },
    "track_justification": "Dynamics is structurally matched: controlled actuated pendulum rollout with recurrent hidden state."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "stage2_bench.py",
    "bench_report.json",
    "stage2_output.txt"
  ],
  "limitations": "Only the small registered dynamics track was tested. The implementation used empirical covariance fitting and barrier/projection rather than a full SDP solution for P and K; larger models, longer horizons, formal robust-invariance certification, speed/FLOP measurements, and alternative ellipsoid estimators were not tested.",
  "system_verdict": "failed",
  "practical_verdict": "harms",
  "mechanism_ok": 0,
  "system_judged": true
}