Given a claim and a corpus, code which studies support it and which contradict it, reading each study's findings field, and report the split without collapsing to a verdict. Emits a direction (supports/contradicts/mixed/not_applicable), the sign of the disputed quantity, and a verbatim justifying quote per study, plus a mechanism-locus flag that surfaces studies whose effect is in the model's output rather than in the person — so a sign-correct count is never mistaken for a claim about human cognition. Use before writing up a result as settled, and whenever a 'most studies find…' statement is about to go in a draft.
