A preregistered ablation study on a 709-page markdown wiki maintained by an LLM tests whether progressive disclosure—using a compact catalog and one-line summaries instead of a monolithic index—reduces costs while maintaining answer quality. The experiment compares four corpus versions across three access conditions: a protocol-constrained agent, a free self-routing agent, and a catalog-preload regime.
- A pilot revealed that capable tool-using agents often bypass the index entirely, inferring page paths directly from questions.
- Answer quality in the retrieval arm was non-inferior to the index baseline within the preregistered margin.
- Costs fell significantly across all regimes: by about a third for the self-routing agent and well over half under catalog-preload.
- Savings resulted from more targeted access, with the retrieval arm citing fewer pages and requiring fewer tool turns.
The study demonstrates that progressive disclosure can lower costs without sacrificing quality, while also serving as a case study in evaluation validity through threat-to-validity discipline.