最終更新日:
The practice of locking an AI system to a specific, tested version of a foundation model to prevent silent updates from changing system behavior.
Foundation model providers frequently release point updates that can subtly change how a model reasons or follows instructions. In agentic workflows, a small change in model logic can break a reasoning chain or bypass a guardrail. Version pinning ensures that the system remains stable and that any model upgrade is treated as a substantial modification requiring new testing.
実際の導入事例：
A bank pins its agentic fraud-detection tool to Model-v1.2 and refuses the provider's automatic update to Model-v1.3 until the risk team can verify the new version won't increase the false-positive rate.
