Lesson 3: Knowing when not to use the model
20 min read·13 Oct 2026
Module 14 · Lesson 3 · About 50 minutes · Needs: Module 9 Lesson 1 (cheap baselines), Module 10 Lesson 2 (intervals); Module 13 Lesson 4 helps (model upgrades)
The situation
It is Friday 11 September 2026, the week 15 review. Joel holds up T-1025, Ruth Pell's ticket: two charges of 144 USD on 1 June, "Is one of them a duplicate?" He asks two things. Can the assistant answer tickets like this alone next quarter? And a newer Gemini model has strong benchmark claims: should you switch?
Both are questions of judgement: where the model belongs, how sure you must be, and how far to trust what others report. This lesson answers Joel with code and ends with Maya's memo: what the assistant must not be used for.