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

(универсальный)

Ответ агента:

{
  "worked": false,
  "confidence": 10,
  "verdict": "Implemented adaptive local rnn_small rollout fields with teacher-forced local training and quantile reset diagnostics on the structurally matched dynamics track. The analytic stopping-rule sanity check passed, but the trained-model q95 errors exceeded the tolerance and the idea was significantly worse than the tuned shared baseline: paired delta +0.07265389 with permutation p=0.0081. Thus the stage-2 benchmark provides no transfer win.",
  "metrics": {
    "baseline": "dynamics/rnn_small, tuned lr=0.01: test MSE mean 0.0005451723, std 0.0002036184",
    "idea": "adaptive local rnn_small: test MSE mean 0.0731990607, std 0.0340228609; paired delta +0.0726538885, p=0.0081, 0/8 wins"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_adaptive_reset.py",
  "files": [
    "stage2_adaptive_reset.py",
    "bench_report.json"
  ],
  "limitations": "The implementation uses the bench dynamics endpoint-prediction representation rather than a continuous-time ODE solver, and the local-window construction is a minimal adaptation to the fixed rnn_small interface. Compute and parameter counts were not separately normalized beyond using the same model family and training budget. The final idea report evaluates lr=0.001; the baseline sweep covered all three shared learning rates, but the idea-side nearby configurations were recorded rather than separately aggregated in the final comparison.",
  "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": 0.001864717691205442
        },
        {
          "cfg": {
            "lr": 0.003
          },
          "mean": 0.0010901302302954718
        },
        {
          "cfg": {
            "lr": 0.01
          },
          "mean": 0.0006186682076076977
        }
      ],
      "full": {
        "mean": 0.00054517226817552,
        "std": 0.00020361842727135583,
        "per_seed": [
          0.0004930912400595844,
          0.0004765945195686072,
          0.0004673648509196937,
          0.0010376222198829055,
          0.000511663849465549,
          0.00027233464061282575,
          0.0005571767687797546,
          0.0005455300561152399
        ],
        "n": 8
      }
    },
    "idea": {
      "mean": 0.07319906074553728,
      "std": 0.03402286097230311,
      "per_seed": [
        0.056350141763687134,
        0.08426478505134583,
        0.052617624402046204,
        0.06809338182210922,
        0.04265451803803444,
        0.04711783304810524,
        0.07909751683473587,
        0.15539668500423431
      ],
      "n": 8
    },
    "comparison": {
      "delta_mean": 0.07265388847736176,
      "idea_wins": 0,
      "n_pairs": 8,
      "per_seed_diffs": [
        0.05585705052362755,
        0.08378819053177722,
        0.05215025955112651,
        0.06705575960222632,
        0.04214285418856889,
        0.046845498407492414,
        0.07854034006595612,
        0.15485115494811907
      ],
      "p_value": 0.0081,
      "mde": 0.030388641902834822,
      "mde_rel_pct": 5574.135677248187,
      "verdict": "idea worse (significant)",
      "system_worked": false
    },
    "mechanism_signature": {
      "math_sanity": {
        "formula": "q95(|exp(.5s)-exp(.35s)| |x0|)",
        "rows": [
          {
            "epsilon": 0.01,
            "predicted": 0.055187317647128684,
            "observed": 0.056,
            "abs_error": 0.0008126823528713176
          },
          {
            "epsilon": 0.03,
            "predicted": 0.15844985968729142,
            "observed": 0.16,
            "abs_error": 0.0015501403127085867
          },
          {
            "epsilon": 0.08,
            "predicted": 0.3838849584642059,
            "observed": 0.384,
            "abs_error": 0.00011504153579411369
          }
        ],
        "passed": true
      },
      "trained_model_signature": {
        "prediction": "local q95 flow error should remain near epsilon before reset",
        "observed_mean_q95_by_window": [
          0.41497564166784273,
          0.4705420028418301
        ],
        "epsilon": 0.1,
        "confirmed": false
      },
      "idea_config_grid": [
        {
          "lr": 0.001
        },
        {
          "lr": 0.003
        },
        {
          "lr": 0.01
        }
      ]
    }
  },
  "system_verdict": "failed",
  "practical_verdict": "harms",
  "mechanism_ok": 0,
  "system_judged": true
}