Technical Setup of "Reasoning Over Recall"

The "Reasoning Over Recall" study investigates whether providing Large Language Models (LLMs) with incomplete legal context via RAG—specifically by omitting controlling precedents—degrades reasoning compared to relying on internal knowledge. Using a "Blacklist Protocol," the experiment evaluates frontier models like GPT-5, DeepSeek R1, and Gemini 2.5 Pro across diverse legal tasks and three distinct judicial personas. The technical framework employs a highly controlled, deterministic pipeline to ensure reproducible and structured results.