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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 10,
  "verdict": "The full registered benchmark protocol was run on terminal_reachability_mdp with eight paired seeds, a tuned baseline sweep, and three shared learning rates for the idea. The exact Bellman/Doob check passed, but the idea was significantly worse on the independent test MSE (paired delta +3.32995e-06, permutation p=0.0081), so worked=false. The learned-message mechanism signature was also not confirmed because terminal violation rate was 5.8%.",
  "metrics": {
    "baseline": "Best lr=0.003; full 8-seed mean test MSE=1.0600038902186526e-08.",
    "idea": "Best shared-grid lr=0.01; full 8-seed mean test MSE=3.3405517725348943e-06; paired delta=+3.329951733632708e-06; p=0.0081; 0/8 paired wins."
  },
  "bench_report": {
    "bench_version": 1,
    "track": "terminal_reachability_mdp",
    "model": "mlp_tiny",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.003,
        "epochs": 24,
        "batch": 64,
        "weight_decay": 0.0
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 24,
            "batch": 64,
            "weight_decay": 0.0
          },
          "mean": 3.224428262171841e-07
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 24,
            "batch": 64,
            "weight_decay": 0.0
          },
          "mean": 1.0160677965664888e-08
        },
        {
          "cfg": {
            "lr": 0.01,
            "epochs": 24,
            "batch": 64,
            "weight_decay": 0.0
          },
          "mean": 1.1143229339261044e-08
        }
      ],
      "full": {
        "mean": 1.0600038902186526e-08,
        "std": 5.2151957363594205e-09,
        "per_seed": [
          3.983920304762023e-09,
          7.282767278127267e-09,
          2.039087654281957e-08,
          8.985147736950694e-09,
          1.0061987687493001e-08,
          1.3581075819502075e-08,
          1.5497944261255725e-08,
          5.016591586581853e-09
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 3.3405517725348943e-06,
      "std": 7.964166334274456e-06,
      "per_seed": [
        4.722510027477256e-08,
        2.3129257442633389e-07,
        5.031787608800187e-08,
        5.248364800536365e-07,
        8.457169542452903e-08,
        7.869610954003292e-07,
        5.988180191707215e-07,
        2.440039133944083e-05
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 3.329951733632708e-06,
      "idea_wins": 0,
      "n_pairs": 8,
      "per_seed_diffs": [
        4.3241179970010535e-08,
        2.2400980714820662e-07,
        2.99269995451823e-08,
        5.158513323166858e-07,
        7.450970773703602e-08,
        7.733800195808271e-07,
        5.833200749094658e-07,
        2.4395374747854248e-05
      ],
      "p_value": 0.0081,
      "mde": 7.120916002585194e-06,
      "mde_rel_pct": 67178.20630937803,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "math_check": {
      "event_probability": 0.7560131375000001,
      "max_normalization_residual": 1.1102230246251565e-16,
      "terminal_violations": 0,
      "horizon": 8
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 24,
          "batch": 64,
          "weight_decay": 0.0
        },
        "full": {
          "mean": 0.0005934698098144509,
          "std": 0.0009439925224443243,
          "per_seed": [
            1.2875242646259721e-05,
            1.535874980618246e-05,
            0.00040430238004773855,
            0.00017721200129017234,
            0.0029863957315683365,
            0.00024860911071300507,
            2.5307252144557424e-05,
            0.0008776980102993548
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 24,
          "batch": 64,
          "weight_decay": 0.0
        },
        "full": {
          "mean": 2.8937354531688442e-05,
          "std": 7.176546462090427e-05,
          "per_seed": [
            4.704144203060423e-07,
            2.0069187485205475e-06,
            3.893134348231797e-08,
            2.345227358091506e-06,
            0.00021874086814932525,
            5.943692826804181e-07,
            7.452385943906847e-07,
            6.5568683567107655e-06
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 24,
          "batch": 64,
          "weight_decay": 0.0
        },
        "full": {
          "mean": 3.3405517725348943e-06,
          "std": 7.964166334274456e-06,
          "per_seed": [
            4.722510027477256e-08,
            2.3129257442633389e-07,
            5.031787608800187e-08,
            5.248364800536365e-07,
            8.457169542452903e-08,
            7.869610954003292e-07,
            5.988180191707215e-07,
            2.440039133944083e-05
          ],
          "n": 8
        }
      }
    ],
    "protocol_note": "baseline and idea share architecture, data, optimizer, epochs, batch and the union of all three learning rates",
    "mechanism_signature": {
      "predicted_terminal_violation_rate": 0.0,
      "observed_terminal_violation_rate": 0.058,
      "predicted_normalization_residual": 0.0,
      "observed_normalization_residual": 0.0,
      "confirmed": false,
      "model_test_mse": 4.722510027477256e-08
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 doob_bench.py",
  "files": [
    "doob_bench.py",
    "bench_report.json",
    "math_check.json"
  ],
  "limitations": "The registered benchmark uses a small finite-state terminal-reachability MDP and the intervention is implemented as exact-feasibility-weighted training visitation, not a full continuous latent trajectory sampler or learned value-network Doob transform. Rejection-sampling wall-clock cost, event-count conditioning, approximate-message KL at neural scale, and continuous pendulum rollout conditioning were not tested. The observed normalization residual is computed after explicit normalization and is not an independent learned-message accuracy test; terminal violations are the substantive mechanism measurement.",
  "system_verdict": "failed",
  "practical_verdict": "harms",
  "mechanism_ok": 0,
  "system_judged": true
}