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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Implemented an end-to-end HOCBF projection around the shared rnn_small dynamics model and ran the full eight-seed paired protocol. Baseline MSE was 0.0107554 and idea MSE was 0.0106944, with paired delta -0.0000609 but permutation p=0.3172, so the improvement is not significant. The shield reduced measured trained-model constraint violations from 9.86% nominally to 1.27% shielded, but the predeclared mechanism confirmation criterion was not met.",
  "metrics": {
    "baseline": "Best lr=0.01; eight-seed mean MSE 0.0107553606, std 0.0007592926.",
    "idea": "lr=0.01, k1=2, k2=2; eight-seed mean MSE 0.0106944281, std 0.0008047806; paired delta=-0.0000609325; p=0.3172; 5/8 wins."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_hocbf.py",
  "files": [
    "bench_hocbf.py",
    "bench_report.json"
  ],
  "limitations": "Only the built-in supervised pendulum dynamics track was tested, not PPO or SAC online policy learning. The barrier used a simplified scalar theta safety condition and approximate known dynamics; disturbances, uncertainty tightening, multiple obstacles, and deployment latency were not tested. Actuator-infeasible states can still violate the barrier.",
  "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": 0.0119797271
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.0114801137
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 0.0108951214
        }
      ],
      "full": {
        "mean": 0.0107553606,
        "std": 0.0007592926,
        "per_seed": [
          0.0110420631,
          0.0103416322,
          0.0106304567,
          0.0115663335,
          0.0096338373,
          0.0099414159,
          0.012083089,
          0.0108040571
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.0106944281,
      "std": 0.0008047806,
      "per_seed": [
        0.0110742887,
        0.0102442931,
        0.0105364083,
        0.0112389727,
        0.009480332,
        0.0099933762,
        0.012290501,
        0.0106972521
      ],
      "n": 8,
      "sweep": [
        {
          "cfg": {
            "lr": 0.01,
            "k1": 2.0,
            "k2": 2.0
          },
          "mean": 0.0107734909
        },
        {
          "cfg": {
            "lr": 0.001,
            "k1": 2.0,
            "k2": 2.0
          },
          "mean": 0.0119274086
        },
        {
          "cfg": {
            "lr": 0.01,
            "k1": 2.0,
            "k2": 2.0
          },
          "mean": 0.0107734909
        }
      ]
    },
    "comparison": {
      "delta_mean": -6.09325e-05,
      "idea_wins": 5,
      "n_pairs": 8,
      "p_value": 0.3172,
      "verdict": "no measurable effect",
      "system_worked": false
    },
    "mechanism_signature": {
      "trained_model_raw_constraint_violation_rate": 0.09859375,
      "trained_model_shielded_constraint_violation_rate": 0.01265625,
      "projection_activation_rate": 0.09859375,
      "confirmed": false
    },
    "custom_track": null
  },
  "system_verdict": "failed",
  "practical_verdict": "no_effect",
  "mechanism_ok": 0,
  "system_judged": true
}