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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "The idea was implemented as an end-to-end spherical RoPE residual transformer and evaluated against a matched standard RoPE transformer on the registered sequence track. The mechanism signature was confirmed at neural-network scale, but the idea significantly worsened test MSE, so it does not qualify as a win.",
  "metrics": {
    "baseline": "Best lr=0.001; full 8-seed test MSE 0.192726 ± 0.022233.",
    "idea": "Best tested lr=0.003, beta=1.0, eta=0.5; full 8-seed test MSE 0.464207 ± 0.069616.",
    "paired_delta": "idea - baseline = +0.271481 MSE; 0/8 idea wins; permutation p=0.0081."
  },
  "bench_report": {
    "bench_version": 1,
    "track": "sequence",
    "model": "transformer_tiny",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.001
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001
          },
          "mean": 0.1927255019545555
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.21616622433066368
        },
        {
          "cfg": {
            "lr": 0.006
          },
          "mean": 0.24436583742499352
        }
      ],
      "full": {
        "mean": 0.1927262246608734,
        "std": 0.022232878429441,
        "per_seed": [
          0.21683096885681152,
          0.1569768339395523,
          0.1901477575302124,
          0.2069464474916458,
          0.19253608584403992,
          0.17629599571228027,
          0.2282254546880722,
          0.17385025322437286
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.46420706808567047,
      "std": 0.0696159664906768,
      "per_seed": [
        0.5837222337722778,
        0.46462294459342957,
        0.42523160576820374,
        0.4688754081726074,
        0.47928518056869507,
        0.4190153479576111,
        0.5342675447463989,
        0.33863627910614014
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 0.27148084342479706,
      "idea_wins": 0,
      "n_pairs": 8,
      "per_seed_diffs": [
        0.3668912649154663,
        0.30764611065387726,
        0.23508384823799133,
        0.2619289606809616,
        0.28674909472465515,
        0.2427193522453308,
        0.3060420900583267,
        0.16478602588176727
      ],
      "p_value": 0.0081,
      "mde": 0.050432977935438064,
      "mde_rel_pct": 26.168196893900337,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "prediction": "Normalized RoPE scores lie in [-1,1], giving softmax floor exp(-2 beta)/n; spherical residual projection preserves unit norm.",
      "predicted_floor": 0.004229227601144147,
      "observed_floor": 0.015373899950645864,
      "max_observed_norm_error": 1.1920928955078125e-07,
      "confirmed": true
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "beta": 1.0,
          "eta": 0.5
        },
        "mean": 0.5600788816809654,
        "std": 0.06620292359906506
      },
      {
        "cfg": {
          "lr": 0.003,
          "beta": 1.0,
          "eta": 0.5
        },
        "mean": 0.46420706808567047,
        "std": 0.06961531160762538
      },
      {
        "cfg": {
          "lr": 0.006,
          "beta": 1.0,
          "eta": 0.5
        },
        "mean": 0.4669463559985161,
        "std": 0.08683987850593247
      }
    ],
    "paired_seed_protocol": {
      "seeds": [
        0,
        1,
        2,
        3,
        4,
        5,
        6,
        7
      ],
      "epochs": 8,
      "n_train": 1200,
      "n_test": 400
    },
    "structural_match": "sequence-level multi-token forecast with attention; same task and end-to-end model family"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py",
  "files": [
    "bench_experiment.py",
    "bench_report.json"
  ],
  "limitations": "Only the registered built-in sequence forecast track was tested, with 8 epochs, 1200 training samples, 400 test samples, two transformer blocks, and three learning-rate settings. The every-K-step spectrum controller, longer sequence lengths, and other built-in tracks were not evaluated.",
  "system_verdict": "partial",
  "practical_verdict": "harms",
  "mechanism_ok": 1,
  "system_judged": true
}