# Эксперимент: Conserved Poisson Feature Noise (#1373)

{
  "worked": false,
  "confidence": 8,
  "verdict": "Built a conserved graph-density noise module with Euler–Maruyama flux, positivity projection, exact mass preservation, and a small digits-token classifier comparison. The numerical check verified nonnegativity and conservation with maximum mass error 2.8e-14 and showed nonzero spatial covariance unlike iid noise. In one matched-variance run, conserved noise reached 95.56% accuracy versus 94.67% for iid noise and 95.33% for no noise, but the small 0.22 percentage-point gain over baseline was not established as reproducible and incurred about 6.5x the training time, so there is no sufficiently clear overall win.",
  "metrics": {
    "baseline": "No noise: test accuracy 0.9533, final train loss 0.0834, 1.59 s",
    "idea": "Conserved Poisson noise: test accuracy 0.9556, final train loss 0.1034, 9.06 s; iid matched control: 0.9467, 0.1354, 1.38 s. Math: maximum mass error 2.84e-14, minimum density 0, density variance 0.650; conserved covariance was -0.0685 at Manhattan distance 1 and remained nonzero through distance 8, while iid covariance was approximately zero."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json",
    "run.log",
    "run_matched.log"
  ],
  "limitations": "Only the sklearn 8x8 digits dataset and a small MLP token model were tested, with one random seed and 15 epochs. No ViT/CIFAR corruption robustness, multi-seed confidence intervals, FLOP-normalized comparison, or larger spatial graphs were evaluated; the density update is a dense NumPy implementation rather than a production sparse kernel."
}