Free-Form AI vs Evidence-Enabled AI: A Structural Comparison of Narrative Drift in Document-Grounded Responses

This white paper presents a structured comparison between two AI response regimes applied to the same documentary artefacts: (1) a free-form large language model (LLM) operating without structural evidence constraints, and (2) an evidence-enabled LLM operating under an explicit documentary traceability framework. Using a controlled corpus of policy and regulatory documents, the study evaluates response behaviour across three dimensions: narrative drift, unsupported inference, and internal contradiction. Each enabled response is restricted to explicitly retrievable excerpts from the source document and is accompanied by verifiable citation spans. The results demonstrate systematic divergence between free-form and evidence-enabled outputs, including hallucinated operational details, numeric inconsistencies, and scope expansion beyond document boundaries. The evidence-enabled regime eliminates unsupported claims by design, though it introduces trade-offs in stylistic fluency and expressive flexibility. This paper documents the methodology, implementation architecture, and empirical results, and provides a reproducible framework for evaluating document-grounded AI systems in governance, regulatory, and clinical contexts.