A customer support org has two candidate custom models in Foundry for intent classification that drive routing rules in multiple apps. You must keep…

The requirements for environment alignment, no auto-upgrades, SLA adherence, stable costs, and fast rollback capabilities point directly to a controlled, version-specific deployment strategy. A "Pinned deployment" explicitly links a specific model version to an environment (e.g., prod uses v1.2), preventing unexpected changes from auto-upgrades and ensuring cost stability. A "rollback runbook" is a predefined procedure to quickly revert to a previous stable version, directly satisfying the requirement to roll back within a single change window. This combination provides the necessary stability, control, and safety for a mission-critical system.

A. A single shared deployment is a poor practice that eliminates environment isolation, making it impossible to test changes without risking the production environment.

C. Fine-tuning without governance directly contradicts the requirement for central enforcement and alignment across dev/test/prod, leading to instability and non-compliance.

D. Auto-upgrading to the latest version is explicitly forbidden by the constraint "can’t allow auto-upgrades" due to the risk of SLA-breaching outages.

1. Microsoft. (2024). MLOps: Model management, deployment, and monitoring with Azure Machine Learning. Microsoft Learn. This document details MLOps best practices, including controlled rollout strategies. The concept of deploying specific, registered model versions aligns with "pinned deployment" to ensure predictability and enable rollbacks. See the "Deploy and manage models with MLOps" section.

2. Microsoft. (2024). Deploy models for inference and prediction in Microsoft Fabric. Microsoft Learn. The documentation on model deployment in Fabric emphasizes versioning ("ML model versions"). Pinning is the act of selecting a specific version for a deployment endpoint, providing the stability required by the scenario.

3. Kreuzberger, D., et al. (2023). Machine Learning Operations (MLOps): Overview, Definition, and Architecture. IEEE Access, vol. 11, pp. 31646-31661. DOI: https://doi.org/10.1109/ACCESS.2023.3262108. This academic review of MLOps highlights the importance of reproducibility and versioning of all assets, including models, as a core principle for reliable production systems. See Section III-B, "Reproducibility."