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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 8,
  "verdict": "Built residual-scenario training with a bootstrap residual buffer, box-constraint slack penalty, and the shared rnn_small model on the controlled-pendulum dynamics track. Across 8 paired seeds, test MSE changed from 0.0014198302 to 0.0013042011, but the paired permutation p-value was 0.6789; both systems had zero measured mean slack, so the claimed safety benefit was not demonstrated.",
  "metrics": {
    "baseline": "mean test MSE 0.0014198302; std 0.0003983911; best config lr=0.01, mu=0.0",
    "idea": "mean test MSE 0.0013042011; std 0.0005283693; best config lr=0.01, mu=0.05",
    "paired_comparison": "delta_mean=-0.0001156291; idea wins 5/8; permutation p=0.6789; verdict=no significant win",
    "mechanism_signature": "baseline residual std=0.03405833, idea residual std=0.03194591, baseline mean slack=0.0, idea mean slack=0.0, confirmed=false"
  },
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "rnn_small",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.01,
        "mu": 0.0
      },
      "sweep": [
        {
          "lr": 0.001,
          "mu": 0.0,
          "mean": 0.0042987791
        },
        {
          "lr": 0.001,
          "mu": 0.05,
          "mean": 0.0042987791
        },
        {
          "lr": 0.001,
          "mu": 0.2,
          "mean": 0.0042987791
        },
        {
          "lr": 0.001,
          "mu": 0.5,
          "mean": 0.0042987791
        },
        {
          "lr": 0.003,
          "mu": 0.0,
          "mean": 0.0034615941
        },
        {
          "lr": 0.003,
          "mu": 0.003,
          "mean": 0.0034615941
        },
        {
          "lr": 0.003,
          "mu": 0.2,
          "mean": 0.0034615941
        },
        {
          "lr": 0.003,
          "mu": 0.5,
          "mean": 0.0034615941
        },
        {
          "lr": 0.01,
          "mu": 0.0,
          "mean": 0.0014865802
        },
        {
          "lr": 0.01,
          "mu": 0.05,
          "mean": 0.0014865802
        },
        {
          "lr": 0.01,
          "mu": 0.2,
          "mean": 0.0014865802
        },
        {
          "lr": 0.01,
          "mu": 0.5,
          "mean": 0.0014865802
        }
      ],
      "full_mean": 0.0014198302
    },
    "idea": {
      "best_cfg": {
        "lr": 0.01,
        "mu": 0.05
      },
      "mean": 0.0013042011,
      "std": 0.0005283693,
      "n": 8
    },
    "comparison": {
      "delta_mean": -0.0001156291,
      "idea_wins": 5,
      "n_pairs": 8,
      "p_value": 0.6789,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "baseline_residual_std": 0.03405833,
      "idea_residual_std": 0.03194591,
      "baseline_mean_slack": 0.0,
      "idea_mean_slack": 0.0,
      "confirmed": false
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py",
  "files": [
    "bench_experiment.py",
    "artifacts/bench_report.json"
  ],
  "limitations": "Only the built-in 400-train/200-test dynamics subset and 12 training epochs were tested. The benchmark measured one-step target MSE and empirical residual slack, not long-horizon rollout constraint violations, regime-conditioned buffers, Gaussian augmentation, or formal scenario-confidence guarantees.",
  "system_verdict": "failed",
  "practical_verdict": "no_effect",
  "mechanism_ok": 0,
  "system_judged": true
}