# Эксперимент: Zero-loss stratum Langevin optimizer (#831)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built the zero-loss stratum Langevin optimizer and verified the underlying Gibbs mechanism numerically: isolated versus manifold zero sets produced observed learning coefficients 0.9999999999999997 and 0.5 versus predictions 1.0 and 0.5, the mass-ratio exponent was 0.5000000000000001 versus 0.5, and width scaling was -0.4999999999999996 versus -0.5. However, in the equal-budget digits MLP experiment the Langevin phase substantially degraded performance rather than improving robustness or generalization, so the proposed optimizer did not produce a practical win in this setup.",
  "metrics": {
    "baseline": "AdamW: train loss 0.0296278, train accuracy 0.998515, test loss 0.0889683, test accuracy 0.971111, perturbed train loss 0.0324021.",
    "idea": "AdamW plus 500 zero-loss Langevin steps: train loss 0.206331, train accuracy 0.922791, test loss 0.258167, test accuracy 0.913333, perturbed train loss 0.217161. Toy predictions matched: manifold lambda 0.5 observed 0.5; point lambda 1.0 observed 1.0; mass-ratio exponent 0.5 observed 0.5000000000000001; width exponent -0.5 observed -0.4999999999999996."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "artifacts/results.json",
    "artifacts/run.log"
  ],
  "limitations": "Only the sklearn digits dataset and one small MLP/configuration were tested; no beta, epsilon, step-size, batch-size, architecture, or multi-seed sweep was performed. The analytic toy verification tests regular zero manifolds and isolated quadratic zeros, not singular varieties or multiplicity logarithmic factors. The implementation uses full-batch updates and isotropic noise on all trainable parameters."
}