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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a local forward-intersection projection variant of the shared rnn_small dynamics model and evaluated it against the standard bench.train_model GRU under paired seeds. The trained-model signature confirmed substantial projection behavior (mean rank 1.75/64 and projection MSE 0.0142), but the idea's test MSE was 0.37190 versus 0.0002042 for the tuned baseline; paired delta was +0.37170 with permutation p=0.0081, a significant loss. The claimed stability-style mechanism appeared, but it destroyed predictive accuracy on this task.",
  "metrics": {
    "baseline": "dynamics/rnn_small; tuned lr=0.01, weight_decay=0.0; 8-seed test MSE mean 0.0002041958, std 0.000109425",
    "idea": "tau=0.15, levels=1, lr=0.01; 8-seed test MSE mean 0.3719000, std 0.1780352; delta +0.3716958, p=0.0081; mean trained-model projection MSE 0.0141751 and mean retained rank 1.75/64"
  },
  "bench_report": {
    "bench_version": 1,
    "track": "dynamics",
    "model": "rnn_small",
    "metric_direction": "lower is better",
    "n_seeds": 8,
    "baseline": {
      "best_cfg": {
        "lr": 0.01,
        "weight_decay": 0.0
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "weight_decay": 0.0
          },
          "mean": 0.0011543919
        },
        {
          "cfg": {
            "lr": 0.003,
            "weight_decay": 0.0
          },
          "mean": 0.0006730457
        },
        {
          "cfg": {
            "lr": 0.01,
            "weight_decay": 0.0
          },
          "mean": 0.0002436721
        },
        {
          "cfg": {
            "lr": 0.003,
            "weight_decay": 0.0001
          },
          "mean": 0.0006697542
        }
      ],
      "full": {
        "mean": 0.0002041958,
        "std": 0.000109425,
        "per_seed": [
          0.0003712306,
          0.0003398617,
          7.11567e-05,
          0.0001924393,
          0.0001771611,
          0.0001133421,
          7.96381e-05,
          0.0002887371
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.3719,
      "std": 0.1780352,
      "per_seed": [
        0.2504995,
        0.3216166,
        0.2485008,
        0.6178864,
        0.1744787,
        0.2290141,
        0.4539971,
        0.6792071
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 0.3716958,
      "idea_wins": 0,
      "n_pairs": 8,
      "per_seed_diffs": [
        0.2501283,
        0.3212767,
        0.2484296,
        0.6176939,
        0.1743015,
        0.2289008,
        0.4539174,
        0.6789183
      ],
      "p_value": 0.0081,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "prediction": "compatible projection reduces unsupported hidden transition energy/rank",
      "mean_projection_mse": 0.0141751,
      "mean_rank": 1.75,
      "raw_spectral_radius_range": [
        1.08086,
        1.11266
      ],
      "confirmed": true
    },
    "custom_track": null
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_runner.py > bench_output.json",
  "files": [
    "bench_runner.py",
    "bench_output.json",
    "article.md",
    "forward_intersection.py",
    "mini_experiment.py"
  ],
  "limitations": "Only the built-in actuated-pendulum dynamics track was tested; no Lorenz or learned encoder/world-model experiment, multi-refresh schedule, exact nullspace intersection, multi-step training loss, or alternative projection tolerances beyond the three tau settings was evaluated. The local intervention used a hidden-state SVD compatibility proxy rather than a full coefficient-space latent dictionary intersection.",
  "system_verdict": "partial",
  "practical_verdict": "harms",
  "mechanism_ok": 1,
  "system_judged": true
}