Text this: Evaluating Architectural Safeguards for Uncertain AI Black-Box Components