Эксперимент: Recurrence-to-Latent Cycling Regularizer (#876)

{ "worked": false, "confidence": 9, "verdict": "Built a PyTorch/NumPy MVP implementing recurrence selection, detached persistence-like weights, the ordered path-minus-shortcut cycle loss, and boundary utilities. The core boundary identity was confirmed exactly, adjacent-edge admission matched the predicted strict r>R_gamma transition, and scaling the weights produced exactly linear loss scaling with zero loss for no features. However, the mini-experiment showed tied accuracy (1.0 vs 1.0) while the idea was substantially slower (10.67s vs 1.15s), so there is no demonstrated practical win.", "metrics": { "baseline": "GRU classification accuracy 1.000; training time 1.153s", "idea": "GRU plus recurrence cycle regularizer accuracy 1.000; training time 10.675s; exact weight-loss ratios [0, 0.25, 0.5, 1, 2]; boundary residual {}" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "The implementation uses a cheap recurrence-vertex proxy rather than a full sparse H0/H1 cubical persistence reduction, does not compare against a Vietoris-Rips baseline, uses a tiny synthetic sine classification task, and the recurrence-threshold sweep is coarse around the analytically predicted nonadjacent transition." }