A document-extraction service calls Claude through a model alias that always resolves to the latest version. After a provider release, the output format shifts slightly and the downstream parser starts rejecting a small share of documents. Nothing changed in your code. Which change best prevents this class of incident?

Answer: C — Pin a specific model version in production and qualify every version change with the regression eval suite

(c) is right: an alias lets behavior change without a deploy on your side; pinning makes a model change a deliberate release gated by the eval suite. (a) hides the symptom, adds cost and still fails when the shift is systematic. (b) is an unfounded generalization and is itself an unqualified model change. (d) is prompt emphasis; it doesn't address the uncontrolled version change.