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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "The registered built-in dynamics track was rerun with 8 paired seeds, a tuned three-point baseline sweep, and the same learning-rate grid for the idea. The matching-controllable recurrent model had higher test MSE than the dense baseline (0.004187 versus 0.003113), paired delta +0.001074, and permutation p=0.3437; therefore there was no significant win. The trained-model mechanism signature was also not confirmed: numerical controllability ranks were 15–18/24 rather than full rank.",
  "metrics": {
    "baseline": "Best lr=0.01, weight_decay=0.0; full 8-seed test MSE mean 0.0031130631 ± 0.0015298698.",
    "idea": "Best lr=0.01, weight_decay=0.0; full 8-seed test MSE mean 0.0041867752 ± 0.0021045838.",
    "delta_mean": 0.0010737121192505583,
    "p_value": 0.3437
  },
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "explicit_tanh_rnn_shared",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.01,
        "weight_decay": 0.0
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "weight_decay": 0.0
          },
          "mean": 0.32027048990130424
        },
        {
          "cfg": {
            "lr": 0.003,
            "weight_decay": 0.0
          },
          "mean": 0.01070007259841077
        },
        {
          "cfg": {
            "lr": 0.01,
            "weight_decay": 0.0
          },
          "mean": 0.0029869703284930438
        }
      ],
      "full": {
        "mean": 0.0031130630668485537,
        "std": 0.0015298698474074236,
        "per_seed": [
          0.002583631779998541,
          0.001341367606073618,
          0.0019145020050927997,
          0.006108379922807217,
          0.005108166951686144,
          0.002453721361234784,
          0.0029577165842056274,
          0.002437018323689699
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.004186775186099112,
      "std": 0.0021045838143091996,
      "per_seed": [
        0.0009666543919593096,
        0.005662341136485338,
        0.0049017248675227165,
        0.0013687764294445515,
        0.007184116635471582,
        0.0038662750739604235,
        0.003245178610086441,
        0.006299134343862534
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 0.0010737121192505583,
      "idea_wins": 2,
      "n_pairs": 8,
      "per_seed_diffs": [
        -0.0016169773880392313,
        0.00432097353041172,
        0.002987222862429917,
        -0.004739603493362665,
        0.0020759496837854385,
        0.0014125537127256393,
        0.0002874620258808136,
        0.0038621160201728344
      ],
      "p_value": 0.3437,
      "mde": 0.002543128344355023,
      "mde_rel_pct": 81.69215623792383,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "prediction": "row-saturating matching gives full-rank trained controllability and improves long-horizon input access",
      "trained_model_measurements": {
        "ranks": [
          16,
          17,
          17,
          15,
          15,
          18,
          17,
          16
        ],
        "min_singular_range": [
          1.5868783161841502e-24,
          6.442173137209217e-16
        ],
        "max_singular_range": [
          0.7894802689552307,
          1.0363144874572754
        ]
      },
      "confirmed": false
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 matching_bench.py",
  "files": [
    "matching_bench.py",
    "bench_report.json",
    "bench_run.log",
    "bench_run_rerun.log"
  ],
  "limitations": "Only the registered built-in dynamics pendulum track was tested. No multi-context Gramian penalty, alternative masks, larger hidden sizes, or longer training budgets were evaluated; numerical rank is tolerance-dependent.",
  "system_verdict": "failed",
  "practical_verdict": "inconclusive",
  "mechanism_ok": 0,
  "system_judged": true
}