A free model earns a place in a review path only after it fails the right cases on purpose. The gate that matters is not a pass rate on happy-path examples; it is a small, hostile fixture set where the correct answer is a refusal, a flag, or an explicit statement of missing context. This article describes that gate as a repeatable procedure, including a runner script, a fixture file, a decision table, and the point at which the whole exercise stops being worth the effort.

A pairing session on this problem tends to produce the same three questions before anyone writes code. Which inputs must never produce a confident answer? Who reads the output when the model flags something? And what happens to the pipeline on the day the free tier is unavailable? The method below answers those questions with artifacts rather than opinions, which is why it survives review.

The three questions a reviewer asks first

What must the model refuse?…