A housing authority runs prompt evaluations that must target the exact foundation-model revision used in staging. What practice enables that?

Versioning foundation-model deployments so prompt evaluations can pin the exact model revision is the correct practice for a housing authority that must evaluate prompts against the staging FM revision. Prompt quality results are only meaningful when the evaluation targets a known model version; versioned deployments let GenAIOps pin eval jobs and staging apps to that revision instead of only pinning the exact wording revision in source control, which leaves the model revision unpinned. That reproducibility is required before housing assistants promote prompt or model changes.

Versioning prompt templates in source control so evals pin the exact wording revision fails because pinning wording doesn't ensure the evaluation targets the exact foundation-model revision used in staging. Aliasing foundation-model endpoints so prompt evals route to the latest staging build fails because routing to the latest build doesn't pin the evaluation to the exact staging revision; the eval still runs against whatever build is latest at that time, not the known revision. Standardizing decoding parameters so prompt evals compare outputs under identical settings fails because identical settings don't guarantee the evaluation targets the exact foundation-model revision; they only control response consistency, not model versioning.

Exam caveat: versioning/pinning is about reproducibility of evals and deployments, not about pinning the exact wording revision in source control. If the stem says evals must target the exact foundation-model revision used in staging, answer with versioned deployments and pinning.

Operational check: publish versioned FM deployments in staging, configure the housing prompt evaluation suite to pin that revision, and block production promotion unless eval results cite the same pinned version.