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

(универсальный)

Ответ агента:

{
  "worked": true,
  "confidence": 9,
  "verdict": "Built and ran a structurally matched custom 1D Bloch/PDE benchmark because no built-in track contained reciprocal-lattice gauge structure. With identical mlp_tiny systems and an equal three-point learning-rate sweep, canonicalized q plus the exact gauge phase achieved significantly lower test MSE than raw-q conditioning. The trained-model mechanism signature also confirmed the predicted reciprocal-shift covariance quantitatively.",
  "metrics": {
    "baseline": "Raw-q mlp_tiny: tuned lr=0.006, 25 epochs; full 8-seed test MSE mean 0.3917703, std 0.0135264.",
    "idea": "Canonicalized q plus exact gauge phase: lr=0.006, 25 epochs; full 8-seed test MSE mean 0.0205089, std 0.0010668; paired delta -0.3712614, 8/8 wins, permutation p=0.0081."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "bloch_track.py",
    "stage2_bench.py",
    "bench_report.json"
  ],
  "limitations": "The custom task is a fast synthetic 1D Bloch-factor surrogate rather than a discretized heterogeneous Maxwell/Helmholtz solver or full Fourier neural operator. Covariance-loss augmentation, physical band-structure quantities, larger models, speed/FLOP tradeoffs, and real material-profile data were not tested.",
  "bench_report": {
    "bench_version": 1,
    "track": "bloch_gauge_pde",
    "model": "mlp_tiny",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.006,
        "epochs": 25,
        "weight_decay": 0.0
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 25,
            "weight_decay": 0.0
          },
          "mean": 0.4257675111293793
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 25,
            "weight_decay": 0.0
          },
          "mean": 0.40371501445770264
        },
        {
          "cfg": {
            "lr": 0.006,
            "epochs": 25,
            "weight_decay": 0.0
          },
          "mean": 0.3903501480817795
        }
      ],
      "full": {
        "mean": 0.39177029952406883,
        "std": 0.013526435694561944,
        "per_seed": [
          0.38321977853775024,
          0.3987618088722229,
          0.379114031791687,
          0.40030497312545776,
          0.37645071744918823,
          0.3769666254520416,
          0.4070354998111725,
          0.4123089611530304
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.02050885744392872,
      "std": 0.0010667659553666602,
      "per_seed": [
        0.02256021834909916,
        0.02063133381307125,
        0.018858235329389572,
        0.02092020958662033,
        0.02036798931658268,
        0.021304797381162643,
        0.01952664740383625,
        0.019901428371667862
      ],
      "n": 8,
      "best_cfg": {
        "lr": 0.006,
        "epochs": 25,
        "weight_decay": 0.0
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 25,
            "weight_decay": 0.0
          },
          "mean": 0.02632583351805806
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 25,
            "weight_decay": 0.0
          },
          "mean": 0.02197333751246333
        },
        {
          "cfg": {
            "lr": 0.006,
            "epochs": 25,
            "weight_decay": 0.0
          },
          "mean": 0.02074249926954508
        }
      ]
    },
    "comparison": {
      "delta_mean": -0.3712614420801401,
      "idea_wins": 8,
      "n_pairs": 8,
      "per_seed_diffs": [
        -0.3606595601886511,
        -0.37813047505915165,
        -0.36025579646229744,
        -0.37938476353883743,
        -0.35608272813260555,
        -0.355661828070879,
        -0.38750885240733624,
        -0.39240753278136253
      ],
      "p_value": 0.0081,
      "mde": 0.012368184909651097,
      "mde_rel_pct": 3.15699912032031,
      "verdict": "idea better (significant)",
      "system_worked": true
    },
    "mechanism_signature": {
      "prediction": "canonicalized model has much smaller trained-model reciprocal-shift gauge discrepancy",
      "rows": {
        "baseline": {
          "shift_m": 1,
          "observed_gauge_rmse": 0.8507512807846069,
          "relative_rmse": 0.9375168128658017
        },
        "idea": {
          "shift_m": 1,
          "observed_gauge_rmse": 5.5333178039518316e-08,
          "relative_rmse": 5.384122013910342e-08
        }
      },
      "confirmed": true
    },
    "custom_track": {
      "name": "bloch_gauge_pde",
      "file": "bloch_track.py",
      "domain": "pde"
    }
  },
  "system_verdict": "worked",
  "practical_verdict": "helps",
  "mechanism_ok": 1,
  "system_judged": true
}