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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 10,
  "verdict": "The projector-splitting LoRA optimizer was run on registered built-in track tabular with eight paired seeds. QR orthogonality was confirmed on trained models, but the method was significantly worse than tuned Adam on test MSE.",
  "metrics": {
    "baseline": "Adam mean MSE 8.84202093.",
    "idea": "Projector splitting mean MSE 155.12437153.",
    "comparison": "delta_mean +146.28235060; p_value 0.0081; idea_wins 0/8."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "stage2_bench.py",
    "bench_report.json"
  ],
  "limitations": "Only tabular mlp_tiny, rank-4 adapter, 400/200 samples, and 18 epochs were tested.",
  "bench_report": {
    "bench_version": 1,
    "track": "tabular",
    "model": "mlp_tiny",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.03,
        "betas": [
          0.9,
          0.999
        ]
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.003,
            "betas": [
              0.9,
              0.999
            ]
          },
          "mean": 193.9741
        },
        {
          "cfg": {
            "lr": 0.003,
            "betas": [
              0.9,
              0.99
            ]
          },
          "mean": 197.7391
        },
        {
          "cfg": {
            "lr": 0.01,
            "betas": [
              0.9,
              0.999
            ]
          },
          "mean": 17.7281
        },
        {
          "cfg": {
            "lr": 0.01,
            "betas": [
              0.9,
              0.99
            ]
          },
          "mean": 17.6246
        },
        {
          "cfg": {
            "lr": 0.03,
            "betas": [
              0.9,
              0.999
            ]
          },
          "mean": 9.1818
        },
        {
          "cfg": {
            "lr": 0.03,
            "betas": [
              0.9,
              0.99
            ]
          },
          "mean": 9.2693
        }
      ],
      "full": {
        "mean": 8.84202093,
        "std": 1.07864748,
        "per_seed": [
          10.2105341,
          7.16484642,
          10.15797043,
          9.19402313,
          7.74287367,
          9.73996449,
          8.46479702,
          8.06115818
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 155.12437153,
      "std": 25.43355486,
      "per_seed": [
        171.24125671,
        187.00245667,
        195.04309082,
        146.14537048,
        144.15423584,
        147.81997681,
        114.5154953,
        135.0730896
      ],
      "n": 8,
      "tested_lr_candidates": [
        0.003,
        0.01,
        0.03
      ]
    },
    "comparison": {
      "delta_mean": 146.2823506,
      "idea_wins": 0,
      "n_pairs": 8,
      "p_value": 0.0081,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "confirmed": true,
      "max_u_orth_error": 6.055e-07,
      "max_v_orth_error": 5.147e-07,
      "max_factor_norm": 18.9931
    }
  },
  "system_verdict": "partial",
  "practical_verdict": "harms",
  "mechanism_ok": 1,
  "system_judged": true
}