Introduction
This is a practical guide for a team that wants to adopt Aegis as the place where the quality of their LLM and agent workflows is measured, tracked, and enforced. It walks the whole arc: getting your data in, choosing what to measure, baselining in development, wiring evaluation into automation (CI/CD, trace ingestion, monitoring), gating on results, and iterating over time. It is deliberately specific about what the platform does for you and what it does not, so you know which parts you build yourself.
The core loop never changes. Every check in Aegis reduces to one shape:
your data (input / prompt / context / output / golden_answer) → a metric scores it (0–100) → that score is compared to a threshold → is_gte_threshold tells you pass or fail, and is_success tells you whether the metric even ran.
Everything below is about feeding that loop reliably, at the right moments, and acting on what comes out.