Use a different-lineage Critic and inject Drafter reasoning

Develop and evaluate a GANDR configuration that uses a Critic from a different model lineage and systematically investigates injection of Drafter reasoning into the Critic’s context, in order to determine whether these changes improve auditing beyond the shared-backbone, separate-context design.

Background

GANDR separates the Drafter and Critic contexts so that the Critic sees the question, retrieved passages, and final draft but not the Drafter’s chain of thought. Both agents nevertheless share backbone weights, leaving possible weight-level self-preference. The paper explicitly identifies two unexplored directions: replacing the Critic with a model from a different lineage and studying whether providing the Drafter’s reasoning to the Critic changes audit quality.