The Deterministic-First Cascade
Run cheap predictable rules first; hand only the exceptions to the AI.
Rules first. AI for the exceptions.
A deterministic-first cascade: cheap, predictable rules handle the bulk, and the LLM only catches what the rules miss.
The trap
Treating the rules as decoration. In a cascade the rules do the bulk of the work; the model only catches what they miss.
Not to be confused with
Not Rules Before Guesses: that card chooses between a rule and AI for a task; the cascade puts both in one pipeline - rules first, the model as the catch.
When to use Designing an AI pipeline that stays cheap and predictable at volume.