Evaluation, Guardrails & Observability
Offline fixtures, calibrated judges, red-teaming, and tracing — evaluation is where model-risk governance meets AI engineering.
Agents fail silently, so evaluation is a production system, not an afterthought. Offline fixture tests replay fixed inputs with expected outputs; live verification checks real runs against source documents. LLM-as-judge is seductive and biased, since judges favor long, confident, self-similar answers, so calibrate judges against human labels and keep deterministic checks for numbers and citations. Red-teaming probes for prompt injection through retrieved documents, a real attack surface in finance. Tracing and cost dashboards make behavior observable. This discipline is where model-risk governance of inventory, validation, and monitoring meets AI engineering, and regulators increasingly expect it.