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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 9,
  "verdict": "Implemented a cyclic-Adam periodic Lyapunov guard around the bench rnn_small model on the structurally matched dynamics track. Across 8 paired seeds, the guard was worse: baseline MSE 4.3824e-05 versus idea MSE 5.2815e-05, paired delta +8.9906e-06, permutation p=0.1569, with 3/8 idea wins. The empirical local update gains were near 1.0 rather than the target 0.95, so there is no demonstrated neural-network training win.",
  "metrics": {
    "baseline": "Constant-rate Adam, best lr=0.01; 8-seed mean dynamics MSE 4.3824038e-05, std 8.7441e-06",
    "idea": "Cyclic Adam with periodic guard, best lr=0.01 and target gamma=0.95; 8-seed mean dynamics MSE 5.2814643e-05, std 1.6132e-05; signature-run final scale 0.9025067"
  },
  "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": 0.001069260761141777
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.00048170724767260253
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 4.429114142112667e-05
        }
      ],
      "full": {
        "mean": 4.3824037675221916e-05,
        "std": 8.74405174104391e-06,
        "per_seed": [
          4.058532795170322e-05,
          5.9273366787238047e-05,
          4.1370050894329324e-05,
          3.593582005123608e-05,
          3.42507082677912e-05,
          3.743398701772094e-05,
          4.52842541562859e-05,
          5.645878627547063e-05
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 5.28146429132903e-05,
      "std": 1.613161621260221e-05,
      "per_seed": [
        5.2862775191897526e-05,
        5.086131568532437e-05,
        4.221191557007842e-05,
        8.595507097197697e-05,
        3.3954747777897865e-05,
        3.525929059833288e-05,
        5.4426076530944556e-05,
        6.698595097986981e-05
      ],
      "n": 8,
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "target": 0.95
          },
          "mean": 0.0010968303249683231
        },
        {
          "cfg": {
            "lr": 0.003,
            "target": 0.95
          },
          "mean": 0.0005092574647278525
        },
        {
          "cfg": {
            "lr": 0.01,
            "target": 0.95
          },
          "mean": 5.797276935481932e-05
        }
      ]
    },
    "comparison": {
      "delta_mean": 8.990605238068383e-06,
      "idea_wins": 3,
      "n_pairs": 8,
      "per_seed_diffs": [
        1.2277447240194306e-05,
        -8.412051101913676e-06,
        8.418646757490933e-07,
        5.001925092074089e-05,
        -2.959604898933321e-07,
        -2.174696419388056e-06,
        9.141822374658659e-06,
        1.0527164704399183e-05
      ],
      "p_value": 0.1569,
      "mde": 1.5079593057768236e-05,
      "mde_rel_pct": 34.40941058311984,
      "verdict": "no significant win",
      "system_worked": false
    },
    "mechanism_signature": {
      "quantity": "local cyclic update gain / periodic gamma",
      "prediction": "guard reduces estimated periodic gain to <= target when unguarded estimate exceeds it",
      "predicted_gamma": [
        0.9999957978598845,
        0.9999967813478916
      ],
      "observed_phase_gains": [
        0.9999812841415405,
        0.9999892711639404,
        0.9999985098838806,
        0.9999967813491821
      ],
      "guard_final_scale": 0.9025066973024747,
      "confirmed": false
    },
    "protocol_notes": {
      "track_match": "dynamics structurally matches stability/control/Lyapunov",
      "epochs": 8,
      "batch": 64,
      "period": 4,
      "shared_lr_grid": [
        0.001,
        0.003,
        0.01
      ],
      "baseline_method": "constant-rate Adam",
      "idea_method": "cyclic Adam with empirical periodic Lyapunov rate backoff"
    }
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py",
  "files": [
    "stage2_bench.py",
    "bench_report.json"
  ],
  "limitations": "Only the built-in dynamics track, rnn_small model, 8 epochs, batch size 64, and three shared learning rates were tested. The practical guard used a low-rank random-probe approximation rather than an SDP or full parameter Jacobian; wall-clock speed and larger models were not evaluated.",
  "system_verdict": "failed",
  "practical_verdict": "inconclusive",
  "mechanism_ok": 0,
  "system_judged": true
}