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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Implemented the Bayesian logit smoother with bursty-mask marginalization as an end-to-end system on a custom persistent latent-sequence classification track. On eight paired seeds, the mechanism prediction was confirmed: trained-model belief distance after five missing steps was approximately 0.677187, matching |lambda_A|^5. The idea nevertheless performed significantly worse than the tuned baseline on test NLL, so there is no benchmark win.",
  "metrics": {
    "baseline": "Mean test NLL 0.311573, std 0.008470, per-seed [0.312010, 0.304609, 0.314550, 0.328771, 0.303857, 0.299804, 0.312833, 0.316153].",
    "idea": "Mean test NLL 0.389429, std 0.020531, per-seed [0.376915, 0.375296, 0.386862, 0.436110, 0.362536, 0.386056, 0.393229, 0.398430].",
    "paired_comparison": "Idea minus baseline delta +0.077856; 0/8 paired wins; permutation p=0.0081; verdict idea worse (significant)."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py",
  "files": [
    "bench_experiment.py",
    "bursty_latent_sequence.py",
    "bench_report.json",
    "bench_run.txt"
  ],
  "limitations": "The custom track is synthetic rather than a real event dataset, and uses a small MLP frame encoder rather than the built-in transformer. Only dropout rate 0.30 and mean burst length 5 were evaluated; learned transition matrices, other burst lengths, calibration, runtime, and FLOPs were not tested.",
  "bench_report": {
    "bench_version": 1,
    "track": "custom:bursty_latent_sequence",
    "model": "mlp_tiny",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.006,
        "epochs": 12,
        "a": 0.95
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 12,
            "a": 0.8
          },
          "mean": 0.559369
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 12,
            "a": 0.9
          },
          "mean": 0.329
        },
        {
          "cfg": {
            "lr": 0.006,
            "epochs": 12,
            "a": 0.95
          },
          "mean": 0.314985
        }
      ],
      "full": {
        "mean": 0.311573,
        "std": 0.00847,
        "per_seed": [
          0.31201,
          0.304609,
          0.31455,
          0.328771,
          0.303857,
          0.299804,
          0.312833,
          0.316153
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.389429,
      "std": 0.020531,
      "per_seed": [
        0.376915,
        0.375296,
        0.386862,
        0.43611,
        0.362536,
        0.386056,
        0.393229,
        0.39843
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 0.077856,
      "idea_wins": 0,
      "n_pairs": 8,
      "p_value": 0.0081,
      "mde": 0.012576,
      "mde_rel_pct": 4.0362,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "confirmed": true,
      "prediction": "missing emissions preserve Markov contraction over q=5 steps",
      "predicted_decay": 0.677187,
      "observed_decay": "approximately 0.677187 on trained-model emissions",
      "q": 5
    },
    "custom_track": {
      "name": "bursty_latent_sequence",
      "file": "bursty_latent_sequence.py",
      "domain": "sequence"
    }
  },
  "system_verdict": "partial",
  "practical_verdict": "harms",
  "mechanism_ok": 1,
  "system_judged": true
}