Capstone: Defend a Model Selection Decision

Outcome. You can produce a decision memo that selects a model system for a real use case and defends it with requirements, eval results, architecture, controls, economics, routing, monitoring, and migration.

The capstone joins the course’s mechanisms into one operational decision. Select a real use case small enough to evaluate: bilingual ticket triage, document extraction, research assistance, internal search, or another bounded workflow. Define what the model receives, what evidence and tools it may use, what it can change, and who is harmed by failure.

Your submission needs seven artifacts: requirements card; system/data-flow diagram; representative eval with English and Khmer when relevant; candidate bake-off; risk and control table; routing/fallback/approval policy; and version/monitor/migration plan. Include raw failures, not only the winning examples. Record current documentation and date because model capabilities and prices age.

A defensible decision may be “do not deploy yet.” It may combine a small classifier, retrieval, deterministic checks, a stronger fallback, and human approval rather than choose one model for everything. The standard is not technical complexity; it is whether another competent team can inspect the evidence, rerun the evaluation, challenge assumptions, and operate the system safely.

Mental model. Model selection is a controlled claim: under these requirements, data, tests, controls, and current conditions, this system is the best supported choice.

Evidence trail — reviewed 23 July 2026. Use the primary sources throughout this course, current official candidate documentation, NIST AI 600-1 (https://doi.org/10.6028/NIST.AI.600-1), and OWASP LLM risks (https://genai.owasp.org/llmrisk/).