# Эксперимент: Conditional-Flow Nested Sampling for Neural Energy Landscapes (#1319)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built an online affine conditional-flow nested-sampling MVP with exact prior/flow importance weighting, weighted resampling, and numerical checks for the Beta(N,1) shrinkage law. The math checks succeeded: empirical E[log t] differed from -1/N by 0.024%, and importance weighting recovered constrained-prior moments with RMSE 0.0275. However, the claimed efficiency win was absent on the 8-D eight-well toy: the flow required 128 score evaluations per replacement versus 2.57 for prior rejection, despite 50.0% proposal acceptance and mean ESS 42.4.",
  "metrics": {
    "baseline": "308 score evaluations, 2.57 evaluations/replacement, mean acceptance 0.644",
    "idea": "15,360 score evaluations, 128 evaluations/replacement, mean candidate acceptance 0.500, mean ESS 42.4"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json",
    "README.md"
  ],
  "limitations": "Only a small 8-D toy landscape, N=64, and 120 replacements were tested. The flow was a diagonal affine Gaussian rather than a full conditional RealNVP or spline flow; no slice/MCMC baseline, evidence-error comparison, GPU implementation, or larger neural energy model was evaluated."
}