# Эксперимент: Thermodynamic Two-State Expert Gate (#1411)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a GRU switching-dynamics benchmark with a thermodynamic two-expert gate, smoothness regularization, and CPU fallback after the shared GPU produced a cuDNN allocation error. The numerical check reproduced the mixture-variance and susceptibility peaks at p=0.5, with finite-difference susceptibility 0.67557 versus the bound 0.67568. However, the learned gate stayed near 0.503 with near-zero regime correlation, and forecasting did not improve consistently, so the promised learned-regime effect was not observed.",
  "metrics": {
    "baseline": "MSE rho=0.1/0.5/0.9: 0.030829/0.033351/0.036821; parameters: 1401",
    "idea": "MSE rho=0.1/0.5/0.9: 0.030626/0.033319/0.036916; parameters: 1443; mean gate approximately 0.503; gate-regime correlation 0.008-0.014; susceptibility peak p=0.5"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "Only a small scalar switching AR system was tested, not Lorenz dynamics, longer-horizon rollouts, an unconstrained MoE control, multiple random seeds, or a calibrated pseudo-label objective. The benchmark used CPU after CUDA/cuDNN failed and the idea model had about 3% more parameters than baseline."
}