Definition
Plain language
A hierarchical checklist of what counts as a good answer to a task, used to grade an AI's work piece by piece.
As stated in the literature
QUEST's central primitive: a hierarchical decomposition of task requirements into binary leaf checks (some marked critical), aggregated upward into branch and root scores, serving both as synthesis target and as reward signal for objective and open-ended deep-research tasks.
Also called: rubric trees
Why it matters: A tree of binary checks turns vague qualities like "good research report" into a mechanical score that's still aligned with what a careful human would look for, making open-ended training tractable.
For example, the root "good market analysis" branches into "identifies competitors" and "estimates market size," each branching further into binary checks like "names at least three competitors."
Heard on the show
“They call it a rubric tree.”Episode 082 — Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick