# Эксперимент: Pointwise complexity-gated inference (#1167)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built and tested a dropout MLP prototype with empirical chaining estimates and adaptive sample allocation. Sparse points had substantially larger chaining Phi (9.30 versus 1.49), but the proposed score was negatively correlated with actual stochastic error (Spearman -0.896), and adaptive inference did not improve accuracy or reduce compute versus fixed sampling.",
  "metrics": {
    "baseline": "Fixed K=30; accuracy 0.9278; mean samples 30.00.",
    "idea": "Adaptive mean K=30.37; accuracy 0.9278; M-error Spearman -0.8955; calibrated failure rate 0.10.",
    "math_sanity": "Passed: sparse-cluster Phi 9.3029 > dense-cluster Phi 1.4866."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 run_experiment.py",
  "files": [
    "run_experiment.py",
    "results.json"
  ],
  "limitations": "Tested only a small two-moons dropout MLP, with one seed and no CIFAR-10, transformer, diffusion, or repeated-seed confidence intervals. The metric and calibration were lightweight MVP approximations."
}