MRRA: An Auditable Reasoning Control Architecture for Role-Specialized LLM Ensembles

This technical report introduces MRRA (Marukoshiki Recursive Reasoning Architecture), an auditable reasoning-control architecture for role-specialized LLM ensembles. Unlike multi-agent debate frameworks optimized for consensus, convergence, or benchmark accuracy, MRRA focuses on decision traceability by combining problem reframing, triangulated role-based evaluation, structured interference analysis, assumption tracking, confidence capping, and conditional revision. MRRA is implemented as a state-machine-based control framework with three core layers: Problem Reframer, Triangulated Evaluator, and Revision Controller. The framework uses an Interference Layer to classify cross-role outputs into Agreement, Conflict, Complement, and Absence / Absent Warning, and maintains an Assumption Registry to preserve decision-relevant assumptions, unresolved risks, and minority warnings. The report includes two case studies: a trading strategy evaluation and a web access-control decision. Both cases demonstrate how MRRA converts multi-LLM reasoning into auditable decision evidence by surfacing critical assumptions, preserving absent warnings, activating confidence caps, and producing conditional recommendations rather than forcing premature consensus.