The simplest architecture that works

Six Larkspur Cloud questions, each already run through several architectures by an earlier module’s lab. Set the data you have, how many model calls you can wait for and what a wrong answer costs. The rules pick the cheapest rung that answers, and the ladder says why each other rung lost.

Question

Data shape

A wrong answer costs

Latency budget: model calls you can wait for, one after another

Picked
Model calls
Retrievals
Checks the evidence

The ladder for this question

Architecture	Calls	Retrievals	What the lab saw	Here

Where the counts come from

Every count is one an earlier lab printed or a notebook’s code fixes, on that lab’s own handbook and question: M10 and M23 (TF-IDF over 12 chunks), M16 (BM25 and all-MiniLM-L6-v2 over 19), M28 (two rewrites allowed), M29 (k = 3), M35 (stub models; tool calls left out, so its retrievals are not counted), M42 and M43 (small fixtures). A rung marked not counted has no course measurement for this question. Calls include the answer call; the notebooks make them one after another.