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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "The contraction-regularized random recurrent layer did not improve the matched dynamics benchmark. Its mean test MSE was 0.0001558971 versus 0.0001305076 for the identical unregularized candidate-map baseline, with paired delta +0.0000253895 and permutation p=0.07125. The trained-model mechanism signature was not confirmed because predicted log expected gain -0.0057538 differed substantially from observed hidden-distance slope -0.3882254.",
  "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": 0.0038340488099493086
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.000746978163078893
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 9.810126357479021e-05
        }
      ],
      "full": {
        "mean": 0.0001305075584241422,
        "std": 5.472800825829195e-05,
        "per_seed": [
          0.00010314843530068174,
          6.60803125356324e-05,
          0.00012244189565535635,
          0.00010073441080749035,
          9.771589247975498e-05,
          0.00018152409757021815,
          0.0001220697449753061,
          0.00025034567806869745
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.00015589705117236008,
      "std": 7.121662686835463e-05,
      "per_seed": [
        0.00011286333028692752,
        5.95987839915324e-05,
        0.00010499265044927597,
        0.00014471430040430278,
        0.00017019343795254827,
        0.00019453300046734512,
        0.00014643977920059115,
        0.00031384112662635744
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 2.538949274821789e-05,
      "idea_wins": 2,
      "n_pairs": 8,
      "per_seed_diffs": [
        9.714894986245781e-06,
        -6.4815285441000015e-06,
        -1.7449245206080377e-05,
        4.397988619681243e-05,
        7.247754547279328e-05,
        1.3008902897126973e-05,
        2.4370034225285053e-05,
        6.349544855765985e-05
      ],
      "p_value": 0.07125,
      "mde": 2.693898995137407e-05,
      "mde_rel_pct": 20.64170863102341,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "predicted_vs_observed": {
        "predicted_log_expected_gain": -0.005753826349973679,
        "observed_log_distance_slope": -0.3882254362106326,
        "relative_slope_error": 66.47245187333293,
        "mean_route_probabilities": [
          0.5307331085205078,
          0.4692668914794922
        ],
        "candidate_spectral_gains": [
          0.8478590250015259,
          1.1598429679870605
        ],
        "test_mse": 0.00011286333028692752,
        "confirmed": false
      },
      "track_justification": "Dynamics is the built-in structural match for recurrent stability and control.",
      "protocol": {
        "epochs": 12,
        "n_train": 1200,
        "n_test": 400,
        "lr_union": [
          0.001,
          0.003,
          0.01
        ],
        "idea_lambda": 0.05,
        "target_rho": 0.95
      }
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_run.py",
  "files": [
    "bench_run.py",
    "bench_report.json",
    "baseline_model.pt",
    "idea_model.pt"
  ],
  "limitations": "Only the registered built-in dynamics track was tested. Sequential MNIST, adding-problem training, GRU comparison, longer training, alternative contraction targets and penalty strengths, sampled routing, and broader model-size sweeps were not tested.",
  "system_verdict": "failed",
  "practical_verdict": "inconclusive",
  "mechanism_ok": 0,
  "system_judged": true
}