Good catch, and half of this is already there. A model's "easy" rating can't send work to the cheap model by itself: a deterministic floor raises it to hard when the profiler sees unit issues, mixed date layouts, or grouping plus rescaling. And easy only gets the small model on attempt 1; any retry escalates to the strong model. What it doesn't do is learn across files. An easy subtask that needed escalation on file N starts on the small model again on file N+1. Keying that on a column profile (dtype, the shapes seen, the source-name hints) and starting on the strong model after repeated escalations is a cheap, measurable fix. The eval set can tell us whether it pays off. Adding it to the roadmap, thanks.