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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Implemented and reran the matched dynamics benchmark on the registered built-in dynamics track. The hyperplane-gated GRU achieved mean test MSE 1.94363e-05 versus the tuned baseline GRU's 6.93655e-06; paired delta was +1.24998e-05 with permutation p=0.02205, yielding the canonical verdict idea worse (significant). The analytical boundary-offset check passed within 0.56%, but the trained-model mechanism signature was not confirmed and there was no benchmark win.",
  "metrics": {
    "baseline": "rnn_small, best lr=0.01; sweep means: lr=0.001 -> 2.68613e-04, lr=0.003 -> 2.24300e-05, lr=0.01 -> 7.06617e-06; full 8-seed mean 6.93655e-06, std 1.29379e-06.",
    "idea": "Hyperplane-gated GRU, best lr=0.01 and temperature=8.0; 8-seed mean 1.94363e-05, std 1.14324e-05; paired delta +1.24998e-05, p=0.02205, 1/8 wins."
  },
  "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": 0.0002686133266252
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 2.2429956516134553e-05
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 7.066165494507004e-06
        }
      ],
      "full": {
        "mean": 6.936550448699563e-06,
        "std": 1.2937909337458348e-06,
        "per_seed": [
          6.290569217526354e-06,
          7.89513342169812e-06,
          7.313619789783843e-06,
          6.7653395490197e-06,
          9.311772373621352e-06,
          7.48660886529251e-06,
          5.2838699957646895e-06,
          5.1454903768899385e-06
        ],
        "n": 8
      }
    },
    "idea": {
      "cfg": {
        "lr": 0.01,
        "temperature": 8.0
      },
      "mean": 1.9436320656041062e-05,
      "std": 1.1432441148988835e-05,
      "per_seed": [
        6.745555310772033e-06,
        2.5030771212186664e-05,
        1.1208207070012577e-05,
        3.957320222980343e-05,
        8.240976967499591e-06,
        2.7031721401726827e-05,
        8.923110726755112e-06,
        2.8737020329572267e-05
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 1.2499770207341498e-05,
      "idea_wins": 1,
      "n_pairs": 8,
      "per_seed_diffs": [
        4.549860932456795e-07,
        1.7135637790488545e-05,
        3.894587280228734e-06,
        3.280786268078373e-05,
        -1.0707954061217606e-06,
        1.9545112536434317e-05,
        3.639240730990423e-06,
        2.3591529952682322e-05
      ],
      "p_value": 0.02205,
      "mde": 1.0426180364417562e-05,
      "mde_rel_pct": 150.30785750822616,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "prediction": "hard routing creates a measurable expert-dependent change near a learned hyperplane, while smooth within-region errors should be lower",
      "observed": {
        "gate_fraction": 0.442,
        "mean_prediction_difference_across_regions": 1.4725773334503174,
        "near_surface_mse": 0.054007068276405334,
        "far_surface_mse": 3.5777430534362793,
        "learned_normal": [
          0.8886291980743408,
          -0.026447903364896774,
          0.07192006707191467
        ],
        "learned_offset": 0.11452355235815048
      },
      "confirmed": false,
      "math_check": {
        "claim": "boundary offset MSE scales linearly as |delta| times jump squared / 2",
        "ratios": [
          1.0052,
          1.0034666666666667,
          1.0034285714285713,
          1.0055466666666666
        ],
        "max_abs_ratio_error": 0.005546666666666589
      },
      "intervention": "learned affine hyperplane with two hard-routed GRU experts"
    },
    "protocol_notes": {
      "structural_match": "dynamics: controlled pendulum rollout",
      "paired_seeds": [
        0,
        1,
        2,
        3,
        4,
        5,
        6,
        7
      ],
      "epochs": 15,
      "baseline_lr_union": [
        0.001,
        0.003,
        0.01
      ],
      "idea_lr_union": [
        0.001,
        0.003,
        0.01
      ]
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "stage2_bench.py",
    "bench_report.json"
  ],
  "limitations": "Only the registered controlled-pendulum dynamics track was tested. The idea sweep used one routing temperature (8.0) and three learning rates; multiple hyperplanes/cells, hysteresis annealing, longer rollout horizons, and broader architectures were not tested.",
  "system_verdict": "failed",
  "practical_verdict": "harms",
  "mechanism_ok": 0,
  "system_judged": true
}