A school-district tutoring bot must only serve a foundation-model version that already passed the district evaluation suite. What should operators do?

Pinning the tutoring bot to the model version that passed the district evaluation suite is the correct operator action for a school-district bot that may only serve an evaluated foundation-model revision. Production configuration should reference the approved version explicitly so pinning the tutoring bot to the stable alias that resolved to the evaluated version cannot introduce an unevaluated model into student-facing tutoring when the alias later resolves to a new revision. Pinning closes the loop between evaluation evidence and runtime serving—a core GenAIOps production control for civic education.

Pinning the tutoring bot to the stable alias that resolved to the evaluated version fails because the alias can later resolve to a new revision that has not passed this district's evaluation suite. Pinning the tutoring bot to the evaluated model checkpoint with tokenizer updates allowed fails because allowing tokenizer updates changes the evaluated serving stack and can put unevaluated behavior in front of students. Pinning the tutoring bot evaluation replay to the version that passed the district evaluation fails because it pins what is re-evaluated, not what model version is actually served to students in production.

Exam caveat: pinning evaluated versions pairs with versioned deployments; together they make eval results actionable. If the stem says the bot must only serve a version that already passed evaluation, answer with pin-to-evaluated-version rather than pinning to the stable alias that resolved to the evaluated version.

Operational check: set the tutoring bot's deployment reference to the exact evaluated FM version, disable auto-upgrade to latest, and require a new eval pass plus change approval before any pin update.