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

(универсальный)

Ответ агента:

{
  "worked": true,
  "confidence": 7,
  "verdict": "Implemented and trained a multiplicative quaternion-manifold recurrent predictor on the registered dynamics track using the canonical bench trainer. Across eight paired seeds, the idea reduced test MSE from 0.006651 to 0.004610, with paired delta -0.002041 and permutation p=0.0317, meeting the harness criterion for a significant win. The unit-quaternion constraint was preserved numerically, but the second-order curvature prediction was not independently quantified at neural-network scale.",
  "metrics": {
    "baseline": "tuned lr=0.006; 8-seed mean MSE 0.0066509384, std 0.0034903632",
    "idea": "lr=0.006, alpha=1.0; 8-seed mean MSE 0.0046102794, std 0.0017827840",
    "paired_delta": -0.0020406589,
    "permutation_p_value": 0.0317
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_stage2.py",
  "files": [
    "bench_stage2.py",
    "bench_report.json"
  ],
  "limitations": "The implementation uses a fixed positive-definite covariance rather than learned process noise, and the baseline/idea recurrent internals are not parameter-count matched exactly. The mechanism signature confirms unit-norm retraction but does not provide a trained-model predicted-versus-observed NEES or curvature-spread measurement.",
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "rnn_small",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.006,
        "alpha": 0.5
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.0015,
            "alpha": 0.5
          },
          "mean": 0.07130480092018843
        },
        {
          "cfg": {
            "lr": 0.0015,
            "alpha": 0.7
          },
          "mean": 0.07130480092018843
        },
        {
          "cfg": {
            "lr": 0.0015,
            "alpha": 1.0
          },
          "mean": 0.07130480092018843
        },
        {
          "cfg": {
            "lr": 0.003,
            "alpha": 0.5
          },
          "mean": 0.02189101977273822
        },
        {
          "cfg": {
            "lr": 0.003,
            "alpha": 0.7
          },
          "mean": 0.02189101977273822
        },
        {
          "cfg": {
            "lr": 0.003,
            "alpha": 1.0
          },
          "mean": 0.02189101977273822
        },
        {
          "cfg": {
            "lr": 0.006,
            "alpha": 0.5
          },
          "mean": 0.009577582241035998
        },
        {
          "cfg": {
            "lr": 0.006,
            "alpha": 0.7
          },
          "mean": 0.009577582241035998
        },
        {
          "cfg": {
            "lr": 0.006,
            "alpha": 1.0
          },
          "mean": 0.009577582241035998
        }
      ],
      "full": {
        "mean": 0.006650938361417502,
        "std": 0.0034903632092782065,
        "per_seed": [
          0.011823982931673527,
          0.005238679703325033,
          0.011270472779870033,
          0.009977193549275398,
          0.0038828086107969284,
          0.004712723195552826,
          0.0035874645691365004,
          0.002714181551709771
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.00461027942947112,
      "std": 0.0017827839604746276,
      "per_seed": [
        0.00657200301066041,
        0.0035921495873481035,
        0.005795476958155632,
        0.0077093965373933315,
        0.002257069107145071,
        0.0037548672407865524,
        0.0043830350041389465,
        0.002818237990140915
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": -0.002040658931946382,
      "idea_wins": 6,
      "n_pairs": 8,
      "per_seed_diffs": [
        -0.005251979921013117,
        -0.0016465301159769297,
        -0.005474995821714401,
        -0.0022677970118820667,
        -0.0016257395036518574,
        -0.0009578559547662735,
        0.0007955704350024462,
        0.00010405643843114376
      ],
      "p_value": 0.0317,
      "mde": 0.001904176545133376,
      "mde_rel_pct": 28.630193841212243,
      "verdict": "idea better (significant)",
      "system_worked": true
    },
    "idea_cfg": {
      "lr": 0.006,
      "alpha": 1.0
    },
    "protocol_note": "8 paired seeds; baseline 3 learning rates x 3 alpha-parity entries; idea 3 alpha/lr settings",
    "mechanism_signature": {
      "predicted_unit_quaternion_norm_error": 0.0,
      "observed_unit_quaternion_norm_error": 0.0,
      "predicted_second_order_gain": "moderate tangent spread",
      "observed_model_mechanism": "unit-norm retraction during recurrent propagation",
      "confirmed": true
    }
  },
  "system_verdict": "worked",
  "practical_verdict": "helps",
  "mechanism_ok": 1,
  "system_judged": true
}