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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 10,
  "verdict": "The resolution-aware schedule was implemented and tested on the structurally matched Poisson PDE track with the shared mlp_tiny architecture and canonical train_model path. Increasing training pairs improved trained-model MSE from 0.0100713 at N=6 to 0.00284859 at N=30 and 0.00102006 at N=165, confirming the mechanism. The tuned baseline selected lr=0.006 and N=165, exactly matching the best idea result; paired delta_mean=0 and permutation p=1.0, so there was no significant benchmark win.",
  "metrics": {
    "baseline": "8-seed mean MSE 0.0011491459, std 0.0001594221; best config lr=0.006, n_train=165.",
    "idea": "Best 8-seed result: kappa=1.5, N=165, lr=0.006, mean MSE 0.0011491459, std 0.0001594221.",
    "paired_comparison": "delta_mean=0.0, p_value=1.0, verdict=no measurable effect",
    "mechanism_signature": "confirmed=true; observed MSE decreased from 0.0100713 at N=6 to 0.00284859 at N=30 and 0.00102006 at N=165."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "stage2_bench.py",
    "bench_report.json"
  ],
  "limitations": "The matched PDE track has a fixed 30-point output grid; the resolution audit used nested subsampling at m=10,20,30 rather than independently generated high-resolution PDE grids. FLOPs were not directly counted. The baseline was allowed to use the largest data budget, so the experiment demonstrates a data-quality mechanism but not a savings or equal-budget advantage.",
  "bench_report": {
    "bench_version": 1,
    "track": "poisson_jfb_short_trace",
    "model": "mlp_tiny",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.006,
        "n_train": 165
      },
      "full": {
        "mean": 0.001149145871750079,
        "std": 0.00015942205170709907,
        "per_seed": [
          0.0010200557298958302,
          0.0010734780225902796,
          0.0011366221588104963,
          0.0010112978052347898,
          0.001334082568064332,
          0.0010319320717826486,
          0.001102412585169077,
          0.0014832860324531794
        ]
      }
    },
    "idea": {
      "best_cfg": {
        "lr": 0.006,
        "n_train": 165,
        "kappa": 1.5
      },
      "mean": 0.001149145871750079,
      "std": 0.00015942205170709907,
      "per_seed": [
        0.0010200557298958302,
        0.0010734780225902796,
        0.0011366221588104963,
        0.0010112978052347898,
        0.001334082568064332,
        0.0010319320717826486,
        0.001102412585169077,
        0.0014832860324531794
      ]
    },
    "comparison": {
      "delta_mean": 0.0,
      "p_value": 1.0,
      "verdict": "no measurable effect",
      "system_worked": false
    },
    "mechanism_signature": {
      "claim": "more operator pairs reduce learned error",
      "confirmed": true,
      "observations": [
        {
          "n_train": 6,
          "test_mse": 0.01007129903882742
        },
        {
          "n_train": 30,
          "test_mse": 0.0028485895600169897
        },
        {
          "n_train": 165,
          "test_mse": 0.0010200557298958302
        }
      ]
    },
    "custom_track": {
      "name": "poisson_jfb_short_trace",
      "file": "/home/maxwelhelp/all/math2nn/bench/custom_tracks/poisson_jfb_short_trace.py",
      "domain": "pde"
    }
  },
  "system_verdict": "partial",
  "practical_verdict": "no_effect",
  "mechanism_ok": 1,
  "system_judged": true
}