# Эксперимент: Projector-Gap Trust Region for Shared Updates (#1346)

{
  "worked": false,
  "confidence": 8,
  "verdict": "Built a projector-gap trust-region controller with shadow post-update projectors, adaptive alpha scaling, retry halving, and a finite-difference verification. The finite-difference error decreased linearly with step size and predicted versus actual projector gap matched exactly; at epsilon=0.90 the controller kept measured gaps below the threshold and reduced some instability. However, at the aggressive learning rate where the baseline diverged, the trust-region version also diverged and did not produce a reliable optimization win, so the full claimed stability benefit was not demonstrated.",
  "metrics": {
    "baseline": "At lr=1.2, epsilon=0.90: final MSE 3.06e27 mean over 5 seeds, 98.8 loss spikes, max gap 0.9343.",
    "idea": "At lr=1.2, epsilon=0.90: final MSE 1.99e27, 78.8 loss spikes, max gap 0.8837, mean alpha 0.9484; gap control worked but divergence remained. At lr=1.0 it was stable with final MSE 0.0367 versus baseline 0.0395, but this was a small toy difference."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "The experiment used four heterogeneous linear recurrent matrices and a quadratic parameter-matching objective rather than GRUs on adding/copy-memory tasks. It did not measure wall-clock or FLOPs, use real optimizer state such as Adam, test long horizons or larger modules, or establish that projector gaps causally predict loss spikes. The trust controller used exact shadow projectors for proposed updates rather than a production Jacobian-vector approximation."
}