Module 3 · Building systems · Lesson 7
Jev alongside LLMs
Hybrid architectures where an LLM generates and Jev decides, as a guardrail on inputs and outputs, a filter on retrieved passages, a citation checker, a selector over extracted candidates and the gate in a cheap-first cascade, with the cost and latency arithmetic that shows when each pays.
By the end of this lesson you can
- Place Jev before, around and after an LLM call, and say which decisions belong to each
- Design a guardrail and a RAG passage filter as batteries of atomic questions with thresholds kept in code
- Implement extract-then-select and draft-then-gate workflows that keep generation in the LLM and decisions in Jev
- Compute the cost, error and latency of a cheap-first cascade with a Jev verifier, and explain how the verifier's recall caps it
- Explain what the System One adapter and the TypeSafe agent skill are for, and what Jev is not