7. A worker model and an expert model

Status: accepted, October 2026

Context

Designing a board takes many AI turns: searching parts, writing the design, reading build errors and fixing them. Running the strongest model for all of it is expensive; running a small or local model for all of it gives worse architecture choices and gets stuck on hard build failures.

Decision

An optional expert model is called only at the points where judgement matters most: planning the architecture once the requirements are clear, reviewing each clean build, after three failed builds in a row, and when the worker asks (consult_expert). The worker does every other turn. Usage is counted per model.

Consequences

• A cheap or local worker with a strong expert costs much less than the expert alone.
• The best pairing is an empirical question, which is what the benchmark is for. No pairing is recommended until it has been run.