Эксперимент: Strongly-convex superwind attention (#1245)

{ "worked": false, "confidence": 8, "verdict": "Built a PyTorch MVP for asymmetric Finsler-style attention with diagonal positive H, learned drift b, scalar profile phi, convexity barrier, and Hessian sanity checks. The math check behaved as expected: phi=1+0.25s gave g approximately 1 and positive sampled Hessian minimum 0.694, while a deliberately nonconvex profile produced a negative minimum of -11.70. However, on the matched tiny synthetic next-token task, constrained metric attention did not beat standard dot attention: final loss was 0.01861 versus 0.01768, so the promised practical win was not observed.", "metrics": { "baseline": "Dot attention: final loss 0.01768; last-20 mean 0.01968; loss std 0.00411; max gradient 0.934; attention entropy 0.748.", "idea": "Constrained asymmetric metric: final loss 0.01861; last-20 mean 0.02085; loss std 0.00447; max gradient 0.949; negative-g fraction 0; attention entropy 0.000002." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a tiny synthetic task and one random seed were tested; no WikiText-2, multi-head/deep Transformer, equal-parameter analysis, finite-difference Hessian sweep during training, or multi-seed statistical comparison was performed. The sampled Hessian evidence validates local behavior, not global Finsler strong convexity." }