# Эксперимент: Inflated-Covariance Convex Chance Constraint (#1359)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a compact covariance-envelope sanity check and a reproducible toy representation-training comparison in experiment.py. The numerical theorem check confirmed Sigma_Y-Sigma_X is PSD and, across tested shifted convex affine regions with inflated probability above one half, the smaller-covariance probability was consistently higher. In the training test, the chance-constraint model retained nearly identical clean accuracy to the clean baseline (0.9589 vs 0.9578) but did not improve acceptance (worst case 0.96595 vs 0.96659), so the proposed effect was not observed as a practical win.",
  "metrics": {
    "baseline": "Clean training: clean accuracy 0.9589; small/inflated/isotropic acceptance 0.9673/0.9666/0.9677. Ordinary inflated-noise augmentation: clean accuracy 0.9511; acceptance 0.8338/0.8027/0.8462.",
    "idea": "Soft inflated-covariance chance penalty: clean accuracy 0.9578; small/inflated/isotropic acceptance 0.9668/0.9660/0.9671. The math sanity check observed inflated/smaller probabilities 0.98661/0.99923 for the primary convex rectangle and correct ordering for all three shifted tests."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py"
  ],
  "limitations": "This is a small synthetic binary toy experiment, not MNIST/CIFAR or a larger representation model. Monte Carlo estimates do not include binomial lower confidence bounds, only one seed and one covariance envelope were tested, and the chance objective uses a sigmoid surrogate rather than a differentiable exact indicator."
}