Abstract Automated literature synthesis using large language models has advanced rapidly through retrieval-augmented generation (RAG), improving coverage and citation grounding. However, existing approaches remain predominantly search-first and content-driven, leaving them vulnerable to topic drift, citation instability, and spurious coherence when synthesizing across large or heterogeneous scientific corpora. We introduce Boundary-First Literature Synthesis (BFLS), a lightweight methodological control framework that constrains retrieval and synthesis using structural boundaries prior to citation aggregation. BFLS operates by (i) identifying invariant constraints, symmetry breaks, and failure modes relevant to a query; (ii) applying ratio- and rhythm-based recurrence filters across domains; and (iii) enforcing deliberate collapse tests to retain only synthesis elements that survive boundary stripping and re-expansion. BFLS is model-agnostic and does not require retraining. It is designed as an overlay compatible with existing RAG pipelines, acting before retrieval ranking and after draft synthesis. We argue that boundary-first control improves citation stability, reduces hallucination under perturbation, and enhances cross-domain coherence without increasing retrieval budget. BFLS reframes literature synthesis as a boundary-conditioned inference task rather than a purely relevance-ranked summarization problem, offering a falsifiable, structure-guided complement to current AI-assisted scientific review systems. Keywords literature synthesis retrieval-augmented generation boundary conditions scientific reasoning hallucination control cross-domain synthesis invariant detection model-agnostic methods