Evals, guardrails, and tracing around an LLM document-extraction agent. Strict tool-use boundary, provenance/totals/schema guardrails, a scored regression-gate eval over a golden set, and a structured trace per run. Runs offline in ~5s.