12 minutes. Mission: recommend use with explicit limits.
Write the decision record
Model cardPurpose: decision and users. Data: population, period, exclusions. Evaluation: split, baseline, metrics, uncertainty. Limits: failure modes and groups not validated. Operations: threshold, fallback, monitoring, retraining trigger.
Compare the candidate against the real baseline and capacity constraint. Inspect errors, not only aggregates. Confirm features exist at serving time. Define what happens when inputs are missing or drift. Google’s production rules emphasize simple first models, pipeline integrity, and monitoring.[1]
Use a strict decision
Go: representative evaluation beats baseline by enough to matter, risks have controls, and monitoring exists. Pilot: evidence is promising but impact or data shift needs bounded real-world testing. No-go: leakage, weak value, unstable scores, harmful errors, or no operational owner.
Final retrieval: explain the whole workflow in 60 seconds without scrolling. Start with decision, end with monitoring.
Check retrieval
Win: you can defend a go, pilot, or no-go recommendation without hiding behind one score.