Decomposes an audit question into explicit, checkable sub-claims with acceptance criteria (what
evidence would make each pass/fail) plus concrete search queries.
This is an alternative to the legacy "think in steps" expansion: by forcing the model to name each
sub-claim and how to verify it up front, the agent has a checklist to work and the verify step has
a rubric to grade against, targeting the "we get the question wrong" failure.
Decomposes an audit question into explicit, checkable sub-claims with acceptance criteria (what evidence would make each pass/fail) plus concrete search queries.
This is an alternative to the legacy "think in steps" expansion: by forcing the model to name each sub-claim and how to verify it up front, the agent has a checklist to work and the verify step has a rubric to grade against, targeting the "we get the question wrong" failure.