# Эксперимент: Heavy-Tail Path-Adaptive Optimizer Pool (#576)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a NumPy MVP of restarted diagonal AdaGrad experts with geometric horizons and the centered variance-normalized exponential meta-controller. Two quantitative checks matched the predicted mechanism: displacement had log-log slope 0.537 versus 0.5 for sqrt(H), and meta log-odds had slope 0.530 versus 0.5 with asymptotic ratios approaching 1. However, the dynamic recovery sweep did not show the claimed O(H) dependence (lags were 4, 6, 7, 7, 7, 8 for H=1,2,4,8,16,32), so the required mechanism criterion was not fully met; the exploratory comparison was also not equal-FLOP fair.",
  "metrics": {
    "baseline": "Shifted quadratic fixed AdaGrad: post-shift loss 6.19796, recovery lag 8 steps",
    "idea": "Pool: post-shift loss 0.04840 +/- 0.00161, recovery lag 12.5 steps; the strong loss result is confounded by multiple expert updates and unequal compute"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 optimizer_pool.py",
  "files": [
    "optimizer_pool.py",
    "results.txt",
    "README.md"
  ],
  "limitations": "Only a small scalar shifted-quadratic toy was tested; no neural fine-tuning, CIFAR, GPU, equal-backward-FLOP study, optimizer wall-clock analysis, or systematic heavy-tail noise-order sweep was performed. The restart implementation resets accumulators but does not reset expert parameters from the weighted model."
}