Fact construction as a missing layer in legal AI: a practice-grounded framework for evaluating multi-agent legal consultation

Legal AI research often evaluates large language models (LLMs) on legal reasoning tasks under the assumption that relevant facts are already structured and available. This assumption overlooks a prior and practically consequential stage of legal work: fact construction, namely the transformation of incomplete, fragmented, and interactionally disclosed information into coherent factual narratives. This study conceptualizes fact construction as a distinct evaluation layer in legal AI and introduces a practice-grounded framework for assessing it in AI-assisted legal consultation. Rather than proposing a legal advice system or doctrinal rule-matching architecture, we use a role-separated multi-agent workflow consisting of a fact-constrained Litigant Twin, a structured-questioning LegalAider, a NarrativeAgent for factual synthesis, and an Evaluator for factual-alignment assessment. Using 100 de-identified legal clinic consultation records, we evaluate multiple LLMs under rubric-based assessment, comparing generated factual narratives against original human-authored consultation records in terms of factual accuracy, completeness, and consistency. The results show that LLM-based agents can generate usable factual narratives under constrained conditions, particularly when cases are internally consistent, limited to a single dispute, and supported by clear actor-role relationships. However, cross-model evaluator agreement remains moderate, and outlier analysis reveals recurring structural sources of failure, including anonymization-induced referential ambiguity, third-party reporting, multi-episode case complexity, and numerically dense facts. These findings suggest that reliability in AI-assisted legal consultation depends not only on model capability, but also on interaction design, data representation, institutional record-keeping, and human oversight. By treating fact construction as a distinct analytical layer, this article contributes a practice-grounded framework for evaluating responsible LLM integration in legal settings and highlights the need for factual grounding, contestability, and human-supervised review before downstream legal analysis begins.