Day 1 — Frame
Define the question, success criteria, and what "good enough" looks like. No code yet.
Days 2–3 — Build
Minimal implementation with known shortcuts. The goal is signal, not production quality.
Day 4 — Eval
Run the defined evals. Collect hard numbers.
Day 5 — Report
One-page decision document: findings, confidence level, recommended next step, and what we would do differently.