A permits office wants reproducible Azure Machine Learning training runs that always pin the same historical applications table. What asset should they create?

Creating a versioned Azure ML data asset that references the historical applications dataset is correct because reproducible training requires pinning an exact municipal dataset snapshot, not "whatever file is newest." Azure ML data assets carry versions so permits-office jobs always resolve the same historical applications table path and metadata. That enables lineage, rollback, and consistent comparisons across forecasting experiments for civic permit workloads.

A versioned Azure ML environment asset for training on the historical applications dataset fails reproducibility: it pins container and package dependencies for the training environment, not the historical applications table, so jobs can still train on different underlying data snapshots. A versioned Azure ML model asset trained on the historical applications dataset fails as a training source of truth—model assets version the scoring output, even when trained on that dataset, not the input dataset version, so they cannot bind runs to the same historical applications table. A versioned Azure ML data asset using the latest historical applications dataset fails conceptually because it floats to whatever version is newest rather than pinning an immutable historical snapshot for training runs.

Exam caveat: datastores connect storage; data assets version specific datasets for consumption. When the stem stresses reproducible runs that pin the same table, choose a pinned versioned data asset—not a versioned environment asset, a versioned model asset, or a data asset pointed at the latest historical dataset.

Operational check: register the historical applications table as a versioned data asset, submit two training jobs referencing that explicit version, confirm both resolve identical paths/hashes as configured, and forbid silent overwrites of the pinned underlying files without cutting a new asset version.