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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built the recursive Bellman variance intervention on the structurally matched dynamics track using the shared rnn_small backbone, with a variance head, detached recursive child variance, and between-branch variance term. Across 8 paired seeds, MSE changed from 0.004214 baseline to 0.004030 for the idea, but the paired permutation p-value was 0.9399 and only 2/8 pairs favored the idea, so there is no significant win. The trained-model mechanism signature was not confirmed: mean calibration ratio was 0.00516 rather than near 1, with weak negative q/error correlation.",
  "metrics": {
    "baseline": "Dynamics/rnn_small MSE: mean 0.00421419, std 0.00266828; tuned lr=0.01, epochs=8.",
    "idea": "Dynamics/rnn_small recursive variance training MSE: mean 0.00402988, std 0.00150342; selected lr=0.01, epochs=8; paired delta -0.00018431, p=0.9399, 2/8 wins.",
    "mechanism_signature": "Mean recursive q=0.81054 versus observed squared error 0.00542; calibration ratio mean=0.00516; q/error correlation mean=-0.0987; confirmed=false."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "stage2_bench.py",
    "bench_report.json"
  ],
  "limitations": "The initial larger run timed out after 2400 seconds, so the completed protocol used 400 training samples, 200 test samples, 8 epochs, and four stochastic branches. The recursive branch construction is an approximation using perturbations of the final input state rather than a learned environment transition model; reward/child covariance was not separately estimated. The signature aggregation contains a reporting simplification for mean q/observed error, although per-seed calibration ratios were measured from each trained model. No rollout allocation or compute-efficiency comparison was tested.",
  "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": 8
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 8
          },
          "mean": 0.0228443113155663
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 8
          },
          "mean": 0.006279829074628651
        },
        {
          "cfg": {
            "lr": 0.01,
            "epochs": 8
          },
          "mean": 0.005194205295993015
        }
      ],
      "full": {
        "mean": 0.004214190877974033,
        "std": 0.0026682777355217157,
        "per_seed": [
          0.002182353986427188,
          0.008899358101189137,
          0.0017826670082286,
          0.007912442088127136,
          0.003060318995267153,
          0.0012072835816070437,
          0.0036685036029666662,
          0.005000599659979343
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.004029877163702622,
      "std": 0.0015034226018154422,
      "per_seed": [
        0.0031153184827417135,
        0.0026313886046409607,
        0.006552070379257202,
        0.0028610709123313427,
        0.005238592624664307,
        0.002022449392825365,
        0.004396332893520594,
        0.005421794019639492
      ],
      "n": 8,
      "selected_cfg": {
        "lr": 0.01,
        "epochs": 8
      }
    },
    "comparison": {
      "delta_mean": -0.0001843137142714113,
      "idea_wins": 2,
      "n_pairs": 8,
      "per_seed_diffs": [
        0.0009329644963145256,
        -0.006267969496548176,
        0.004769403371028602,
        -0.0050513711757957935,
        0.002178273629397154,
        0.0008151658112183213,
        0.0007278292905539274,
        0.0004211943596601486
      ],
      "p_value": 0.9399,
      "mde": 0.003068421703670571,
      "mde_rel_pct": 72.81164504693035,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "prediction": "recursive predicted variance should calibrate observed squared error near 1 and track heteroscedasticity",
      "predicted_vs_observed": {
        "mean_q_recursive": 0.8105446696281433,
        "mean_observed_squared_error": 0.005421794019639492,
        "calibration_ratio_mean": 0.005157590081528752,
        "q_error_correlation_mean": -0.09872341295990021
      },
      "confirmed": false
    }
  },
  "system_verdict": "failed",
  "practical_verdict": "inconclusive",
  "mechanism_ok": 0,
  "system_judged": true
}