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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Rank-8 hidden-state projection slightly improved the dynamics test MSE, but the paired permutation test was not significant, so the idea did not produce a protocol-counted win. The trained-model mechanism signature was confirmed: observed effective rank 8.0 matched the predicted rank.",
  "metrics": {
    "baseline": "Dense GRUCell, tuned lr=0.006: test MSE mean 0.002105819425196387 over 8 seeds.",
    "idea": "Rank-8 projected GRUCell, lr=0.006: test MSE mean 0.0020776986493729055; paired delta -2.812077582348138e-05; permutation p=0.15145."
  },
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "rnn_small",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.006
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.0015
          },
          "mean": 0.004987414460629225
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.0034615940821822733
        },
        {
          "cfg": {
            "lr": 0.006
          },
          "mean": 0.0018282315868418664
        }
      ],
      "full": {
        "mean": 0.002105819425196387,
        "std": 0.000389436566226016,
        "per_seed": [
          0.002004395006224513,
          0.001437882543541491,
          0.0015748648438602686,
          0.002295783953741193,
          0.0024020641576498747,
          0.0024898042902350426,
          0.0020851334556937218,
          0.0025566271506249905
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.0020776986493729055,
      "std": 0.00037367759212413065,
      "per_seed": [
        0.0020221038721501827,
        0.001397693995386362,
        0.0016216225922107697,
        0.002313514007255435,
        0.0023468395229429007,
        0.002421593526378274,
        0.0019859536550939083,
        0.0025122680235654116
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": -2.812077582348138e-05,
      "idea_wins": 5,
      "n_pairs": 8,
      "per_seed_diffs": [
        1.770886592566967e-05,
        -4.0188548155128956e-05,
        4.675774835050106e-05,
        1.7730053514242172e-05,
        -5.522463470697403e-05,
        -6.821076385676861e-05,
        -9.917980059981346e-05,
        -4.4359127059578896e-05
      ],
      "p_value": 0.15145,
      "mde": 4.193884781531501e-05,
      "mde_rel_pct": 1.9915690449765808,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "predicted_rank": 8,
      "observed_effective_rank_mean": 8.0,
      "predicted_factor_scalar_ratio": 5.882352941176471,
      "observed_projection_defect_mean": 1.149918068676925e-05,
      "confirmed": true
    },
    "idea_sweep": [
      {
        "rank": 8,
        "lr": 0.0015,
        "mean": 0.0052895165426889434,
        "std": 0.0027286265828909257
      },
      {
        "rank": 8,
        "lr": 0.003,
        "mean": 0.0029902143724029884,
        "std": 0.0018055794197106632
      },
      {
        "rank": 8,
        "lr": 0.006,
        "mean": 0.0020776986493729055,
        "std": 0.00037367759212413065
      },
      {
        "rank": 16,
        "lr": 0.0015,
        "mean": 0.005181030704989098,
        "std": 0.0027043397351147937
      },
      {
        "rank": 16,
        "lr": 0.003,
        "mean": 0.0029607619217131287,
        "std": 0.0017947703875401172
      },
      {
        "rank": 16,
        "lr": 0.006,
        "mean": 0.002110591303789988,
        "std": 0.00038488654224676564
      },
      {
        "rank": 32,
        "lr": 0.0015,
        "mean": 0.0051692417182493955,
        "std": 0.002700498589073299
      },
      {
        "rank": 32,
        "lr": 0.003,
        "mean": 0.002949635687400587,
        "std": 0.0017909695309575413
      },
      {
        "rank": 32,
        "lr": 0.006,
        "mean": 0.0021058720449218526,
        "std": 0.00038964953095754133
      }
    ],
    "protocol_notes": {
      "n_train": 400,
      "n_test": 200,
      "epochs": 12,
      "matched_architecture": "same GRUCell(3,64)+linear head; only hidden-state rank projection differs",
      "baseline_lr_union": [
        0.0015,
        0.003,
        0.006
      ],
      "idea_ranks": [
        8,
        16,
        32
      ],
      "idea_best_cfg": {
        "rank": 8,
        "lr": 0.006
      },
      "baseline_best_lr": 0.006
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_stage2.py",
  "files": [
    "bench_stage2.py",
    "bench_report.json"
  ],
  "limitations": "Direct peak GPU memory, wall-clock speed, reversible QR/projector-splitting updates, checkpointing, and reconstructed adjoint gradients were not measured; the factor-state ratio is a storage proxy.",
  "system_verdict": "partial",
  "practical_verdict": "no_effect",
  "mechanism_ok": 1,
  "system_judged": true
}