# Эксперимент: ESS-Controlled Autoregressive CTMC Sampler (#1329)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built an exact-enumeration binary Ising sampler with collapsed autoregressive proposals, event-driven single-site CTMC refinement, stable normalized ESS, path likelihood accumulation, and generator validation. Generator columns summed to approximately 1e-15 and the ESS implementation matched the direct formula. CTMC refinement reduced empirical KL from 1.24 to 0.52 at T=1 and 0.33 at T=2, but ESS was not reliably improved and path correction reduced it further, so the claimed ESS-controlled sampling win was not demonstrated.",
  "metrics": {
    "baseline": "Collapsed autoregressive proposal: normalized ESS 0.00257, empirical KL 1.2404, ground-state hits 1711/4096; uniform mutation control KL 1.6264 and ground-state hits 52/4096.",
    "idea": "CTMC at T=1: mean jumps 2.72, empirical KL 0.5207, ESS from A0 0.00140, path-corrected ESS 0.000575, ground-state hits 1663/4096. At T=2, KL was 0.3294 and ESS 0.0116, but this was not a consistent ESS improvement."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 ess_ctmc_experiment.py",
  "files": [
    "ess_ctmc_experiment.py"
  ],
  "limitations": "Only a small L=8 binary Ising test was run, using a deliberately collapsed independent proposal rather than a trained neural autoregressive model or learned CTMC rate head. Batch rejection, adaptive horizon training, gradient optimization, autocorrelation, FLOP-matched timing, and larger categorical systems were not tested."
}