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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built and evaluated Positive-Regime Observable ReLU on the structurally matched controlled-pendulum dynamics track. The math check passed exactly, and trained-model behavior confirmed reduced negative preactivations and a smaller ReLU-affine gap. Nevertheless, the idea's test MSE was worse than baseline and the paired permutation test was non-significant.",
  "metrics": {
    "baseline": "lr=0.01, 20 epochs; mean test MSE=0.0001016568, std=0.0000333668.",
    "idea": "lr=0.01, penalty=0.01, 20 epochs; mean test MSE=0.0001038706, std=0.0000340591.",
    "comparison": "paired delta=+0.0000022138, idea wins=4/8, permutation p=0.70145, verdict=no significant win.",
    "mechanism_signature": "Baseline negative-preactivation fraction=0.6035 and affine gap=0.24998; idea fraction=0.5734 and gap=0.16072; confirmed=true.",
    "math_check": "Affine difference=0.0, negative fraction=0.0, perturbation residual=1.03e-15, pass=true."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "stage2_bench.py",
    "bench_report.json",
    "stage2_run.log"
  ],
  "limitations": "Only the built-in dynamics track was tested with 1,200 training and 400 test samples for 20 epochs. The observability determinant/conditioning loss and explicit coordinate-selection readout were not included in the end-to-end intervention; only the positivity penalty was tested. The baseline and idea used small three-configuration sweeps.",
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "rnn_small",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.01,
        "epochs": 20
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 20
          },
          "mean": 0.00029752701448160224
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 20
          },
          "mean": 0.00014188402201398276
        },
        {
          "cfg": {
            "lr": 0.01,
            "epochs": 20
          },
          "mean": 0.00010523172568355221
        }
      ],
      "full": {
        "mean": 0.00010165680305362912,
        "std": 3.336682519215708e-05,
        "per_seed": [
          0.0001742699823807925,
          6.733651389367878e-05,
          7.70786646171473e-05,
          0.00010224174184259027,
          8.755998715059832e-05,
          0.00010097782796947286,
          7.315044058486819e-05,
          0.00013063926598988473
        ],
        "n": 8
      }
    },
    "idea": {
      "best_cfg": {
        "lr": 0.01,
        "epochs": 20,
        "penalty": 0.01
      },
      "mean": 0.00010387055726823746,
      "std": 3.405913194481688e-05,
      "per_seed": [
        0.0001575992937432602,
        8.58209896250628e-05,
        7.609691965626553e-05,
        0.00010509935236768797,
        7.766245835227892e-05,
        0.00012542385957203805,
        5.6170345487771556e-05,
        0.00014709123934153467
      ],
      "n": 8,
      "sweep": [
        {
          "cfg": {
            "lr": 0.01,
            "epochs": 20,
            "penalty": 0.01
          },
          "mean": 0.00010615413884806912
        },
        {
          "cfg": {
            "lr": 0.01,
            "epochs": 20,
            "penalty": 0.1
          },
          "mean": 0.0001076593453035457
        },
        {
          "cfg": {
            "lr": 0.01,
            "epochs": 20,
            "penalty": 1.0
          },
          "mean": 0.000183743754405441
        }
      ]
    },
    "comparison": {
      "delta_mean": 2.2137542146083433e-06,
      "idea_wins": 4,
      "n_pairs": 8,
      "per_seed_diffs": [
        -1.6670688633753294e-05,
        1.848447573138401e-05,
        -9.817449618069761e-07,
        2.857610525097698e-06,
        -9.8975287983194e-06,
        2.44460316025652e-05,
        -1.6980095097096637e-05,
        1.6451973351649947e-05
      ],
      "p_value": 0.70145,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "math_check": {
        "max_affine_difference": 0.0,
        "negative_fraction": 0.0,
        "max_perturbation_residual": 1.0295876722904796e-15,
        "pass": true
      },
      "baseline_observed": {
        "negative_preactivation_fraction": 0.603515625,
        "mean_relu_affine_gap": 0.24997861124575138,
        "finite_difference_output_change": 1.8480932340025902e-08
      },
      "idea_observed": {
        "negative_preactivation_fraction": 0.5734405517578125,
        "mean_relu_affine_gap": 0.1607181280851364,
        "finite_difference_output_change": 1.2689270079135895e-08
      },
      "confirmed": true
    },
    "protocol_notes": {
      "structural_match": "controlled pendulum rollout is a recurrent dynamics task",
      "n_train": 1200,
      "n_test": 400
    }
  },
  "system_verdict": "partial",
  "practical_verdict": "inconclusive",
  "mechanism_ok": 1,
  "system_judged": true
}