Dual LLM

Also written: dual LLM pattern

A pattern, described by Simon Willison in 2023, that uses a quarantined model to read untrusted text and a privileged model to call tools. The privileged model sees only the quarantined model's structured output.