Эксперимент: Graded Levy-area recurrent state (#815)

{ "worked": true, "confidence": 9, "verdict": "Built the graded Levy-area state with exact skew-symmetric area, diagonal quadratic variation, dilation-compatible normalization, and chunk composition. All quantitative mechanism checks matched prediction exactly: reversal area scaled as |a|/2 and canceled under reversal, repeated-pair area accumulated as n/2, dilation ratios were 1.0 for u, A, and Q, and chunking error was 2.7e-15. The small benchmark did not show an accuracy win: Levy and additive pooling both reached 0.512, while Levy was slower and used more parameters.", "metrics": { "baseline": "Additive pooling: test accuracy 0.5117, 82 parameters, 0.268 s; GRU: 0.4844 accuracy, 514 parameters, 1.456 s.", "idea": "Graded Levy state: test accuracy 0.5117, 154 parameters, 2.722 s; reversal sweep exact, linear accumulation exact, dilation ratios 1.0, chunk error 2.7e-15." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a tiny synthetic order-sensitive classification task was tested, not permuted-MNIST, ListOps, or language modeling. The benchmark used full-batch training and did not equalize hidden-state bytes or FLOPs; CUDA cuDNN failed on the shared device, so the reported benchmark used CPU fallback." }