Эксперимент: Schäffer-Covariant Isometric Recurrent Layer (#1174)

{ "worked": true, "confidence": 8, "verdict": "Built a finite Schäffer-lift module with eigendecomposition-based defect operator, covariance residual, numerical checks, and a paired-initialization CUDA mini-task. The core claim was clearly verified: pre-truncation total-state energy drift was 2.21e-12 and the active-subspace one-step energy error was 1.09e-12, while the ordinary hidden state decayed to 1.80e-08 over the same short horizon. However, the lift did not improve task MSE in the paired delayed-signal test, and a random covariance pair had residual 1.87, so the evidence supports structural stability rather than a general accuracy or covariance win.", "metrics": { "baseline": "Short-horizon hidden norm ratio mean 1.80e-08; long-horizon ratio 1.27e-40; paired-task MSE 0.9019917 -> 0.9018074.", "idea": "Short-horizon lifted total-state norm ratio 1.0000000000007; pre-boundary relative energy drift 2.21e-12; paired-task MSE 0.9019917 -> 0.9018074; finite-truncation long-horizon total ratio 1.04e-32 after boundary loss." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 schaffer_experiment.py", "files": [ "schaffer_experiment.py", "results.json" ], "limitations": "Only a small synthetic delayed-signal task was tested, not permuted-MNIST or sequential CIFAR. The exact infinite lift was approximated by finite truncation, covariance was measured but not trained or parameterized, and no equal-FLOP or multi-seed statistical comparison was performed." }