Эксперимент: Overshoot Budget Controller (#1273)

{ "worked": true, "confidence": 8, "verdict": "Built an Overshoot Budget Controller for scalar SGD, plus an exact asymmetric-Huber construction check for Lemma 3.1. Across 2,000 random schedules there were zero bound violations, with minimum observed R/G ratios of 0.9999999999997 and 0.9999999999964; ordering-dependent bounds differed for permutations of the same steps. In the deliberate four-spike quadratic stress test, the controller reduced every 20x proposal to approximately 1.25–1.66, capped maximum loss at 0.32 and reached final loss 1.38e-15, while unclipped SGD reached maximum loss 7.94 and global gradient clipping reached 175.0. This is a clear toy stability signal, not evidence of a general training win.", "metrics": { "baseline": "Unclipped SGD: final loss 0.01994, max loss 7.938, max |x| 3.985, 0/4 spike steps clipped. Gradient-clipped control: final loss 146.33, max loss 174.98, max |x| 18.707, 0/4 proposal steps reduced.", "idea": "Controller: final loss 1.378e-15, max loss 0.320, max |x| 0.8, 4/4 spikes clipped; accepted spike steps [1.248, 1.403, 1.517, 1.662] versus proposal 20.0. Math check: 2,000 trials, zero failures." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 overshoot_controller.py", "files": [ "overshoot_controller.py" ], "limitations": "Only a scalar quadratic/asymmetric-Huber toy experiment was tested; no MLP, Transformer, CIFAR-10, WikiText-2, AdamW integration, wall-clock comparison, hyperparameter sweep, or repeated random training runs were performed. The gradient-clipping baseline uses fixed gradient clipping rather than a separately tuned learning-rate/clipping pair, so the comparison is illustrative rather than a definitive optimizer benchmark." }