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

(универсальный)

Ответ агента:

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built a local private spectral-whitening front-end and evaluated it end-to-end with the standard sequence forecasting Transformer, using the canonical train_model path and an equal learning-rate sweep. The idea achieved lower test MSE than the tuned baseline on all 8 paired seeds: 0.4811 versus 0.5426, delta -0.06145, permutation p=0.0081, so the harness verdict is idea better (significant). The model-derived mechanism signature was not confirmed: input lag-1 correlation increased rather than decreased after whitening, indicating the accuracy gain is real in this setup but not explained by the intended decorrelation prediction.",
  "metrics": {
    "baseline": "Sequence track, transformer_tiny, tuned lr=0.0015, 8 epochs: test MSE mean 0.542589, std 0.050570; per-seed [0.581503, 0.460040, 0.474068, 0.612520, 0.567659, 0.546453, 0.579340, 0.519132].",
    "idea": "Private spectral whitening, alpha=1.0, smooth=5, lr=0.0015, 8 epochs: test MSE mean 0.481137, std 0.036006; all 8 paired seeds won; delta -0.061453, permutation p=0.0081. Mechanism signature: baseline lag-1 correlation 0.4115 versus idea 0.8136, confirmed=false."
  },
  "bench_report": {
    "bench_version": 1,
    "track": "sequence",
    "model": "transformer_tiny",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.0015,
        "epochs": 8
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.0015,
            "epochs": 8
          },
          "mean": 0.5320326462388039
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 8
          },
          "mean": 0.6313047111034393
        },
        {
          "cfg": {
            "lr": 0.006,
            "epochs": 8
          },
          "mean": 0.9853104948997498
        }
      ],
      "full": {
        "mean": 0.5425892136991024,
        "std": 0.050569796260962196,
        "per_seed": [
          0.5815026760101318,
          0.46004006266593933,
          0.474067747592926,
          0.6125200986862183,
          0.5676589012145996,
          0.546452522277832,
          0.5793400406837463,
          0.5191316604614258
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.48113658651709557,
      "std": 0.036005799238620045,
      "per_seed": [
        0.5375266075134277,
        0.45927196741104126,
        0.46048879623413086,
        0.46748945116996765,
        0.4295176863670349,
        0.459766149520874,
        0.5304144024848938,
        0.5046176314353943
      ],
      "n": 8,
      "best_cfg": {
        "lr": 0.0015,
        "alpha": 1.0,
        "smooth": 5
      }
    },
    "comparison": {
      "delta_mean": -0.061452627182006836,
      "idea_wins": 8,
      "n_pairs": 8,
      "per_seed_diffs": [
        -0.0439760684967041,
        -0.0007680952548980713,
        -0.013578951358795166,
        -0.1450306475162506,
        -0.1381412148475647,
        -0.08668637275695801,
        -0.04892563819885254,
        -0.014514029026031494
      ],
      "p_value": 0.0081,
      "mde": 0.0470281472067518,
      "mde_rel_pct": 8.667357555108287,
      "verdict": "idea better (significant)",
      "system_worked": true
    },
    "mechanism_signature": {
      "input_lag1_corr": 0.8136333167430572,
      "pred_shift_sensitivity": 0.5143288373947144,
      "predicted_direction": "lower input correlation and lower shift sensitivity",
      "observed": "trained-model test predictions",
      "confirmed": false,
      "baseline_input_lag1_corr": 0.41153473666857415,
      "observed_input_lag_reduction": -0.402098580074483
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_private_whitening.py",
  "files": [
    "bench_private_whitening.py",
    "bench_report.json"
  ],
  "limitations": "Only the built-in sequence track was tested; no vision, dynamics, language, or custom track was run. The local privacy mechanism is the provisional clipped Gaussian-noise MVP rather than a formally audited optimal LDP estimator. The idea sweep varied learning rate but held alpha=1 and smoothing=5 fixed, and the mechanism signature was measured on one captured seed and did not confirm the predicted decorrelation direction.",
  "system_verdict": "worked",
  "practical_verdict": "helps",
  "mechanism_ok": 0,
  "system_judged": true
}