# Эксперимент: Conservative Parallel-Edge Decomposition (#1121)

{
  "worked": true,
  "confidence": 9,
  "verdict": "Built a conservative parallel-edge architecture with per-edge channel MLPs whose summed flows are aggregated by a signed incidence matrix. All three quantitative predictions were confirmed: internal residual stayed at floating-point scale across K=1..32, leakage residual was linear in amplitude with fitted slope 0.964352 versus predicted 0.964352, and rollout mass drift was zero for conservative updates but linear in steps for leakage. In the toy graph-flow task, the conservative model improved test RMSE from 0.180608 to 0.160494 while reducing mean mass residual from 0.0635853 to 8.10e-8.",
  "metrics": {
    "baseline": "Unconstrained edge model: test q RMSE 0.180608; mean absolute mass residual 0.0635853; 170 parameters.",
    "idea": "Conservative two-channel edge model: test q RMSE 0.160494; mean absolute mass residual 8.10e-8; 162 parameters. Math sweeps: channel residual <=7.1e-15, leakage fitted/predicted slope 0.964352/0.964352, and 80-step conservative drift 1.71e-14 versus predicted leakage drift 0.56."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 conservative_edge_experiment.py",
  "files": [
    "conservative_edge_experiment.py",
    "results.json"
  ],
  "limitations": "This is a synthetic one-step graph-flow experiment rather than a real battery or energy-system dataset. The exact conservation result is an algebraic property of incidence aggregation, and the baseline comparison uses one fixed graph and one random seed; long-horizon learned forecasting, external-flow modeling, and broader seed sweeps were not tested."
}