# Эксперимент: Phase-Polytope Robust Neural Dynamics (#1416)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built a small output-space phase-polytope experiment with four shared-feature heads, centroid prediction, exact worst-vertex robust training, and an independent-error variance sanity check. The variance check matched Var(mean)=sigma^2/M closely, and robust training reduced worst-vertex RMSE substantially at a modest centroid-RMSE cost. This is a promising but narrow signal rather than proof of improved dynamics modeling.",
  "metrics": {
    "baseline": "Centroid-only training: centroid RMSE 0.02325, max vertex RMSE 0.36145, mean vertex RMSE 0.24270, prediction spread 0.05713.",
    "idea": "Centroid-plus-worst-vertex training: centroid RMSE 0.02942, max vertex RMSE 0.03301, mean vertex RMSE 0.03196, prediction spread 0.000665; independent variance ratios for M=1,2,4,8,16 were 0.995, 1.003, 0.999, 0.998, 1.000."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 phase_polytope_experiment.py",
  "files": [
    "phase_polytope_experiment.py",
    "phase_polytope_results.json"
  ],
  "limitations": "Only one-step prediction on a synthetic oscillator was tested; no multi-step rollout stability, calibration metric, FLOP/speed comparison, cyclic alignment procedure, sampled interior convex combinations, or independent deep-ensemble baseline was evaluated. The phase offsets were synthetic and fixed, and the single-head control has a different training objective, so the comparison is not a definitive end-to-end benchmark."
}