# Эксперимент: Policy-Guided Terminal Trust Region for Optimizers (#1234)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a differentiable H=2 policy-guided terminal trust-region optimizer prototype and a numerical contraction sanity check. The sanity check matched the claimed behavior: exact q=0.75, fallback deviations decayed geometrically, the bound held, and an r=0.5 terminal set remained feasible. However, on the fixed two-class toy MLP, trust-region training reached loss 0.6815 versus Adam's 0.4796 at equal 180 steps, and versus 0.2670 for Adam at the same 720 gradient evaluations; it therefore did not demonstrate an optimization or speedup win.",
  "metrics": {
    "baseline": "Adam, 180 steps / 180 gradient evaluations: final full loss 0.479636; mean last-20 minibatch loss 0.4821457.",
    "idea": "H=2 trust-region, 180 steps / 720 gradient evaluations: final full loss 0.681474; mean last-20 minibatch loss 0.690973; median post-projection terminal deviation 0.01889, maximum 0.03495 (radius 0.035). Equal-budget Adam: final full loss 0.266996 at 720 gradient evaluations. Contraction check: q=0.75 and geometric bound/terminal feasibility both true."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "Only a tiny synthetic two-dimensional classification task and one fixed seed were tested; no MNIST/CIFAR benchmark, rho/radius sweep, multi-seed statistics, Adam moment-state rollout, FLOP-normalized wall-clock study, or empirical neural-network q estimation was performed."
}