Эксперимент: Bennett-whitened gradient trust region (#931)

{ "worked": false, "confidence": 9, "verdict": "Built a diagonal, delayed-covariance Bennett-whitened trust-region controller and numerically verified the root equation to machine precision. In 400 bounded-noise paths, only 0.25% crossed the boundary, so the controller downscaled just 0.13% of training steps; matched toy regression showed nearly identical but slightly worse final loss (0.1735 vs 0.1723) and loss variance (0.001890 vs 0.001856), with no stability win.", "metrics": { "baseline": "SGD: final loss 0.172262, best loss 0.082479, last-50 loss variance 0.001856, 768 steps, no divergence", "idea": "Bennett controller: final loss 0.173519, best loss 0.083908, last-50 loss variance 0.001890, 768 steps, 0.13% downscaled steps, median z/r 0.00382, no divergence; math root error 0 and boundary-crossing path fraction 0.0025" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a small synthetic regression MLP was tested; no CIFAR-10, ViT, AdamW, gradient-clipping control, or multi-seed evaluation was run. The covariance is a diagonal one-step-delayed proxy rather than the theorem's true predictable covariance, and CUDA runtime fallback was not stress-tested." }