# Эксперимент: Fractional Boundary-Factored Neural Solver (#1274)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built and tested the fractional boundary-factored MLP in experiment.py. The numerical check reproduced slopes -0.2495 versus expected -0.25 for the raw derivative and 0.3500 versus expected 0.35 for the weighted quotient gradient. In the matched toy run, boundary-strip relative L2 improved from 0.4213 to 0.1598, but global relative L2 worsened from 0.0779 to 0.0977, so the effect is a real boundary-accuracy win rather than an overall accuracy or speed win.",
  "metrics": {
    "baseline": "global relative L2=0.077865; boundary-strip relative L2=0.421280; final loss=0.001269; runtime=12.17s",
    "idea": "global relative L2=0.097693; boundary-strip relative L2=0.159800; final loss=0.002812; runtime=16.94s"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "The test is a 1D manufactured supervised fitting proxy, not a genuine fractional PDE quadrature/PINN residual experiment. Only one seed, one order, one regularization weight, and one architecture were tested; no higher-dimensional domain or true fractional operator was evaluated."
}