# Эксперимент: Distributed E-Value Prediction Sets (#398)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built and ran a fixed-seed distributed e-value ensemble MVP with calibration normalizers, uncertainty-attenuated convex fusion, and prediction-set inference. Fused empirical mean e-values matched the convex weighted prediction to at most 1.1e-16 across all kappa values, coverage stayed above 0.9995 at alpha=0.1, and increasing kappa reduced average set size from 3.57 to 2.84 while increasing singleton rate from 0.052 to 0.161. The mechanism manifested, but the idea did not beat the strongest-expert baseline on set size at alpha=0.1.",
  "metrics": {
    "baseline": "Strongest expert: point accuracy 0.9659, e-value coverage 0.99994, average set size 2.92814. Unweighted fusion: coverage 0.99954, average set size 3.56554, singleton rate 0.05208.",
    "idea": "Uncertainty-weighted fusion at kappa=2: coverage 0.99996, rejection 0.00004 versus Markov bound 0.09926, average set size 3.06882, singleton rate 0.11758, point accuracy 0.99810. Kappa 0 to 5 reduced set size from 3.56554 to 2.83658."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 distributed_evalue_experiment.py",
  "files": [
    "distributed_evalue_experiment.py",
    "results.json",
    "run_output.txt"
  ],
  "limitations": "This is a synthetic toy ensemble, not CIFAR-10, federated learning, or a graph neural network. Calibration validity is empirical with a 1% safety factor rather than a formal finite-sample construction; uncertainty weights are fixed externally, and learned data-dependent weights, distribution shift, repeated random seeds, runtime, communication cost, and larger neural models were not tested."
}