This slide is course-added. It applies the chapter’s planning factors to a decision the case now faces: how to reuse a language model.

How the slide uses the five objectives

Benefits and costs: the purchase or API price excludes evaluation, monitoring and migration. Planning factors: schedule, lifetime, expertise, criticality and platform decide between the options, as on the six-factor slide. Frameworks and inversion of control: agent frameworks call your code and tools, so the chapter’s ‘the framework calls you’ applies. Configuration: prompts and model settings are configuration and must be versioned with the software. COTS: a hosted model is a COTS component whose internal behaviour you cannot inspect or fix.

Model deprecation. Suppliers retire model versions on their own schedule. A replacement can accept the same requests and still behave differently, so a version change is a supplier interface change in the chapter’s sense.

Evaluation set. A fixed collection of realistic cases with expected outcomes, drawn from the intended workflow. Re-running it on each version is the model equivalent of testing the product instance.

Data terms. Whether requests are stored, for how long and where, is set by the supplier’s terms. For personal data this is a requirement to check before reuse, not after.

The case is course-created. Optional further reading: Sculley et al. (2015), Hidden Technical Debt in Machine Learning Systems.