Эксперимент: Differentially Passive Neural Blocks (#1150)

{ "worked": true, "confidence": 8, "verdict": "Built and numerically verified the discrete residual contraction condition MᵀM−(1−α)I⪯0, including a direct contractive-versus-expansive sanity check. In the controlled experiment, the constrained block had no sampled violations (maximum −0.048), trajectory ratios around 0.847, and 20-step distance ratio 0.036, while the baseline had maximum violation 0.548, mean trajectory ratio 1.088, and distance growth to 5.34. The robustness effect is real in this toy setting, but it incurred a substantial task-MSE cost (0.0345 versus 0.000109), so this is evidence for stability rather than an overall accuracy win.", "metrics": { "baseline": "task MSE 0.000109; max sampled LMI violation 0.548; mean trajectory ratio 1.088; 20-step distance ratio 5.338", "idea": "task MSE 0.0345; max sampled LMI violation −0.048; mean trajectory ratio 0.847; 20-step distance ratio 0.0364" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 safe_experiment.py", "files": [ "experiment.py", "results.json", "safe_experiment.py", "safe_results.json" ], "limitations": "Only a tiny 2D residual regression block was tested, not MNIST or depths 10/50/100. The certification is structural/empirical over sampled inputs rather than a formal dense-domain or interval certificate; the dynamic continuous-time LMI and general learned P parameterization were not tested." }