What should teams do on model-update day?

Direct answer: On model-update day, do not silently trust the new model with critical workflows. Freeze what matters, re-run a golden prompt, note what changed, and name an owner before you widen use.

What this means at work

Updates can improve answers and break habits in the same afternoon. Tone shifts. Tool behavior shifts. A workflow that was “good enough” yesterday can invent new failure modes today.

This is a Boundaries habit: enthusiasm for better models is fine. Unexamined cutovers are not.

Do this next

1. List the AI workflows that touch customers, money, or compliance.
2. Freeze those workflows for a short window or require extra review.
3. Re-run one saved golden prompt per workflow. Capture diffs in a shared note.
4. Assign an owner to say ship, wait, or roll back the habit (not the vendor).

Skip the all-hands hype thread until the golden prompts have been re-run. Silence for a morning beats a broken customer workflow.

FAQ

Do we need a full eval suite?

Start with one golden prompt per critical job. Theater benchmarks can wait. Real diffs cannot.

What if nothing looks different?

Still record “no material diff” with a date and owner. Silence without a note is not a check.

See also

• Pulse: Score the demo before the workflow
• Field Guide: Team norms
• Field Guide: Evaluate without theater