The model-harness-context pipeline is gape-limited, like a snake whose jaws and stomach cap the size of prey it can swallow. Capability gains beyond that limit aren’t lost, but they go unused as capability overhang. A gain of a different shape is another matter. It’s less like prey that’s too big and more like prey the snake isn’t built to eat in the first place. That’s not capability overhang but paradigm overhang, where the pipeline has to change form, not just open wider.