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

(универсальный)

Ответ агента:

{
  "worked": true,
  "confidence": 8,
  "verdict": "Implemented a custom graph track because no built-in track contains graph topology. The degree-2 fermionic GNN achieved mean test error 0.0 across 8 paired seeds versus 0.20875 for the swept baseline, with paired delta -0.20875 and permutation p=0.0312, satisfying the benchmark significant-win criterion. The trained-model signature confirmed exact zero wedge output for odd incident degrees; this is evidence on the custom cycle/path task, not broad graph-learning superiority.",
  "metrics": {
    "baseline": "GraphNet baseline, best lr=0.003 and 12 epochs: mean test err 0.20875, std 0.20704; per-seed [0.0, 0.29, 0.5, 0.015, 0.5, 0.325, 0.0, 0.04]. Sweep means: lr 0.001=0.30750, lr 0.003=0.20125, lr 0.006=0.21625.",
    "idea": "Degree-2 fermionic GraphNet, lr=0.003 and 12 epochs: mean test err 0.0, std 0.0; all 8 per-seed errors were 0.0. Paired delta=-0.20875, 6 wins and 2 ties, permutation p=0.0312."
  },
  "bench_report": {
    "bench_version": 1,
    "track": "cycle_parity_graph",
    "model": "graphnet",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.003,
        "epochs": 12
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 12
          },
          "mean": 0.3074999889358878
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 12
          },
          "mean": 0.2012499978300184
        },
        {
          "cfg": {
            "lr": 0.006,
            "epochs": 12
          },
          "mean": 0.2162499960977584
        }
      ],
      "full": {
        "mean": 0.20874999731313437,
        "std": 0.20704392538301097,
        "per_seed": [
          0.0,
          0.28999999165534973,
          0.5,
          0.014999999664723873,
          0.5,
          0.32499998807907104,
          0.0,
          0.03999999910593033
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.0,
      "std": 0.0,
      "per_seed": [
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": -0.20874999731313437,
      "idea_wins": 6,
      "n_pairs": 8,
      "per_seed_diffs": [
        0.0,
        -0.28999999165534973,
        -0.5,
        -0.014999999664723873,
        -0.5,
        -0.32499998807907104,
        0.0,
        -0.03999999910593033
      ],
      "p_value": 0.0312,
      "mde": 0.18507366186801333,
      "mde_rel_pct": 88.65804275455608,
      "verdict": "idea better (significant)",
      "system_worked": true
    },
    "mechanism_signature": {
      "math_checks": {
        "odd_degree_max_residual": 0.0,
        "parity_max_error": 0.0,
        "wedge_energy": {
          "predicted": 24,
          "observed": 23.964393615722656,
          "ratio": 0.9985164006551107
        }
      },
      "prediction": "Odd incident degree is annihilated exactly at neural-network scale.",
      "predicted_max_residual": 0.0,
      "observed_max_residual": 0.0,
      "confirmed": true,
      "n_observations": 24
    },
    "custom_track": {
      "name": "cycle_parity_graph",
      "file": "bench_fermionic.py",
      "domain": "graph-nn"
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_fermionic.py",
  "files": [
    "bench_fermionic.py",
    "bench_report.json"
  ],
  "limitations": "Only the custom cycle-versus-path graph classification task was tested; no molecular, social, or irregular real-world graph benchmark was evaluated. The graph size was fixed at 12 nodes, higher exterior degrees and explicit circuit regrouping were not tested, and FLOP/latency comparisons were not measured.",
  "system_verdict": "worked",
  "practical_verdict": "helps",
  "mechanism_ok": 1,
  "system_judged": true
}