This poster addresses a core friction in AI-assisted research: Open Science organizes knowledge through provenance, reproducibility, and licensing (source-first epistemic governance). LLMs, by contrast, scale via probabilistic pattern compression and speed—fast, but with risks: attribution gaps, hallucinations, license blindness, and update drift. Rather than attempting to “bake OS values into model weights,” we propose a complementary orchestration: a curated, provenance-aware knowledge-graph layer mediates between OS resources and LLM generation. We illustrate this with two exemplary workflows: (1) Production — LLMs assist in lifting OS corpora into structured, versioned knowledge graphs; human curation plus versioning and changelogs ensure auditability and rollback. (2) Use — The knowledge-graph layer guides context for the LLM so that outputs come with verifiable source anchors, reducing hallucinations and improving attribution and license-aware reuse. Thus, an apparent contradiction becomes a productive coupling: OS bounds the search space and makes provenance visible; LLMs deliver speed, generative agility and expressive power. This results in a faster orientation with verifiable sources—complementarity over conversion. Keywords: Open Science; Knowledge Graphs; Large Language Models; Provenance; Attribution; Governance; Auditability;