A specialized evaluator rather than a conversationalist — this model is trained to score and rank text outputs rather than generate them. It operates quietly in the background of reinforcement learning pipelines, assigning reward signals to guide other models. Compact at 1.8B parameters, it trades broad capability for focused judgment.