# Эксперимент: Geometrically Attracting Random Recurrent Layer (#1101)

{
  "worked": true,
  "confidence": 8,
  "verdict": "The random recurrent layer, spectral-gain penalty, routing, and toy verification were implemented. Expected-product and time-varying product predictions matched Monte Carlo within 0.28%, and the predicted mixed-gain boundary p=0.50 was observed. The proposed nonlinear mixture showed a negative long-horizon distance slope, but no task-training win was established.",
  "metrics": {
    "baseline": "Fixed tanh RNN, recurrent scale 1.18: observed distance slope -0.0010.",
    "idea": "Candidate scales 0.68 and 1.28, expansive probability 0.35: expected gain 0.89, predicted log gain -0.1165, observed distance slope -0.0283; boundary predicted and observed at p=0.50.",
    "product_checks": "Independent and time-varying product-law maximum relative error 0.27%."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py && /home/maxwelhelp/main/bin/python3 smoke_test.py",
  "files": [
    "experiment.py",
    "random_attracting_rnn.py",
    "smoke_test.py",
    "results.json"
  ],
  "limitations": "Sequential MNIST, adding-problem training, GRU comparison, and learned regularized training were not tested. The nonlinear fixed-map comparison is affected by tanh saturation."
}