Cianplan

Level 5: AI / LLM systems · L5.6 Evals & guardrails

Layered guardrails

Design input/output filtering and permissions outside the model.

11 min · Pro

What you take away

  • Four gates — input, retrieval, tool call, output — and only the tool-call gate changes what is possible; the rest change probabilities.

And 3 more inside the class, in Pro.

How the class is laid out

  1. The problem
  2. Bumpers, and what they do not cover
  3. Error analysis first
  4. Judges
  5. RAG metrics
  6. Prompt injection
  7. Guardrails
  8. Check your understanding