Эксперимент: Adaptive Barrier-Margin Regularization (#410)

{ "worked": false, "confidence": 9, "verdict": "Built and ran an adaptive barrier-margin regularization MVP with baseline, fixed-margin, and adaptive-margin policies plus an inference-time projection filter. The update’s fixed point, convergence factor, and linear scaling with disturbance amplitude matched predictions closely, but the mean-error Robbins–Monro update produced about 0.625 coverage rather than the claimed 0.9 coverage. The mechanism is visible as an adaptive safety/tracking tradeoff, but the full safety claim did not hold, so worked=false.", "metrics": { "baseline": "Across sigma={0.05,0.10,0.20,0.40}: post-filter violations 0.497, intervention 0.700, tracking MSE 0.498.", "idea": "Adaptive epsilon={0.0447,0.0894,0.1787,0.3575}; post-filter violations={0.0003,0.0174,0.0671,0.1223}, intervention=0 in all cases, tracking MSE={0.759,0.835,0.998,1.370}. Predicted epsilon fixed point 0.355556 vs observed 0.355556; predicted convergence factor 0.775 vs observed 0.775000; predicted and observed coverage were approximately 0.6247 and 0.625." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "run_output.txt", "REPORT.md" ], "limitations": "This is a scalar supervised toy experiment, not cart-pole, RL, a learned observer, or a real QP safety filter. Gaussian disturbances were used; quantile/conformal calibration, long-horizon closed-loop dynamics, and formal safety guarantees were not tested." }