Try the Small Model First

The large model answers everything well and costs ten times more; the small model answers most things well. Ask the small one first and escalate only what it gets wrong.

Implement cascade(question, small, large, accept) returning {"answer": str, "escalated": bool}.

• Ask small. If accept(answer) returns True, you are done; never touch large.
• Otherwise ask large and return its answer with escalated: True.

accept is the gate, and it must be cheap and deterministic (a format check, a validator, a regex): a model deciding whether the small model was right saves no call. In these tests the gate checks the answer is a bare number.

The tests script small so each path is reproducible (one script answers acceptably, one with prose the gate rejects); large is real. You are scored on cost, so both models' call counts are checked.