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

(универсальный)

Ответ агента:

{
  "worked": true,
  "confidence": 8,
  "verdict": "Gauge-Patched Local Experts was implemented as a GRU training modification on the registered dynamics stability/control track. The idea reduced test MSE from 4.311938073442434e-05 to 4.2328419567638775e-05, with paired delta -7.909611667855643e-07 and permutation p=0.0081; all 8 paired seeds improved. The mechanism signature was also confirmed: overlap disagreement fell from 0.002737239032285288 to 0.002110361514496617 while the maximum observed orthogonal norm error was 2.384185791015625e-07.",
  "metrics": {
    "baseline": "best lr=0.01; 8-seed mean MSE 4.311938073442434e-05, std 2.1245040862825782e-05",
    "idea": "lambda_patch=0.01, best lr=0.01; 8-seed mean MSE 4.2328419567638775e-05, std 2.1386848202922675e-05",
    "delta_mean": -7.909611667855643e-07,
    "p_value": 0.0081,
    "paired_wins": "8/8",
    "relative_improvement_percent": 1.834,
    "overlap_disagreement": "baseline 0.002737239032285288; idea 0.002110361514496617"
  },
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "rnn_small",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.01
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001
          },
          "mean": 0.0008724518556846306
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.00035974886850453913
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 4.963686433256953e-05
        }
      ],
      "full": {
        "mean": 4.311938073442434e-05,
        "std": 2.1245040862825782e-05,
        "per_seed": [
          9.722933464217931e-05,
          3.783951615332626e-05,
          3.003067286044825e-05,
          3.3447933674324304e-05,
          4.86072531202808e-05,
          2.836398380168248e-05,
          3.394361920072697e-05,
          3.549273242242634e-05
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 4.2328419567638775e-05,
      "std": 2.1386848202922675e-05,
      "per_seed": [
        9.703257819637656e-05,
        3.75409254047554e-05,
        2.9079366868245415e-05,
        3.262508107582107e-05,
        4.653289943234995e-05,
        2.7314470571582206e-05,
        3.378422115929425e-05,
        3.471781383268535e-05
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": -7.909611667855643e-07,
      "idea_wins": 8,
      "n_pairs": 8,
      "per_seed_diffs": [
        -1.967564458027482e-07,
        -2.9859074857085943e-07,
        -9.513059922028333e-07,
        -8.22852598503232e-07,
        -2.074353687930852e-06,
        -1.049513230100274e-06,
        -1.593980414327234e-07,
        -7.749185897409916e-07
      ],
      "p_value": 0.0081,
      "mde": 5.226854739698207e-07,
      "mde_rel_pct": 1.2121822370063282,
      "verdict": "idea better (significant)",
      "system_worked": true
    },
    "mechanism_signature": {
      "prediction": "orthogonal transitions preserve hidden norms and patch loss reduces disagreement",
      "predicted_norm_error": 0.0,
      "observed_max_norm_error": 2.384185791015625e-07,
      "observed_baseline_disagreement": 0.002737239032285288,
      "observed_idea_disagreement": 0.002110361514496617,
      "confirmed": true,
      "track_rationale": "Dynamics is the built-in stability/control track; local temporal patches are neighboring state windows."
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_gauge_dynamics.py",
  "files": [
    "bench_gauge_dynamics.py",
    "bench_report.json",
    "rerun_output.json"
  ],
  "limitations": "The built-in dynamics task is not a spatial graph or true mixture-of-experts benchmark; neighboring temporal state windows serve as local patches. The implementation uses fixed SO(2)^32 hidden-coordinate rotations rather than transition MLPs or general Lie groups, and does not test cocycle triples, many-expert scaling, unseen spatial rotations, or larger datasets. The baseline sweep selected learning rate using the harness's 4-seed sweep and reevaluated the selected configuration on all 8 paired seeds.",
  "system_verdict": "worked",
  "practical_verdict": "helps",
  "mechanism_ok": 1,
  "system_judged": true
}