Эксперимент: Affine-to-Axis Distillation (#1397)

{ "worked": false, "confidence": 8, "verdict": "Built a reproducible Affine-to-Axis Distillation MVP with positive-scale fitting, Hungarian matching, normalized axis loss, affine residual measurement, and a two-hidden-layer PyTorch distillation test. The synthetic check reproduced the intended signal: affine-related representations recovered all axes, recovery degraded smoothly with noise, and unrelated representations scored much worse. The task experiment showed only a small one-run accuracy/loss improvement while learned axis quality was substantially worse than baseline, so the practical effect is not demonstrated.", "metrics": { "baseline": "Digits MLP: validation accuracy 0.9711, validation cross-entropy 0.1295, final axis quality 0.9406, normalized-cost<0.25 fraction 0.9688", "idea": "Digits MLP with logit plus axis loss: validation accuracy 0.9733, validation cross-entropy 0.1250, final axis quality 0.6670, normalized-cost<0.25 fraction 0.2344; synthetic noise-0 axis quality 1.000 and noise-0.4 quality 0.850, versus random control 0.379" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 affine_axis_distill.py", "files": [ "affine_axis_distill.py", "results.json", "article.md", "idea_context.json" ], "limitations": "Only one fixed-seed run was performed on sklearn digits rather than CIFAR-10. Repeated seeds, confidence intervals, calibration, merging quality, assignment stability, depth windows, next-layer rescaling, and full global-affine-map training were not tested." }