Эксперимент: Residual-Scenario Safety Training (#1376)

{ "worked": true, "confidence": 7, "verdict": "Built a frozen affine output-head MVP using bootstrap residual scenarios, box constraints, and an exact nonnegative slack penalty. The slack elimination identity matched brute-force minimization to 2.2e-16. Across five seeds, residual-scenario training reduced held-out disturbance violation rate by 24.0% and mean slack by 27.0%, while nominal clean-target MSE increased 10.6%; this is a promising but modest safety signal, with essentially unchanged p95 error.", "metrics": { "baseline": "5-seed mean: nominal MSE 0.0001694, p95 absolute error 0.19770, violation rate 0.003919, mean slack 0.0001557", "idea": "5-seed mean: nominal MSE 0.0001873 (+10.6%), p95 absolute error 0.19784 (+0.07%), violation rate 0.0029775 (-24.0%), mean slack 0.0001137 (-27.0%)" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 multi_seed.py", "files": [ "experiment.py", "multi_seed.py" ], "limitations": "This is a scalar one-step toy with a frozen affine head, iid Gaussian held-out disturbances, and no multi-step dynamics, regime-conditioned sampling, Gaussian-augmentation control, calibration analysis, or formal scenario-confidence validation. The evaluated violation rate is low and p99 slack is zero, so the magnitude and generality of the safety improvement remain uncertain." }