The Integration Bottleneck in AI-Assisted Research: Candidate Artifacts, Expert Attention, and the Post-Generation Problem

AI-assisted research is often discussed as a problem of generation: whether systems can produce papers, proofs, code, reviews, summaries, hypotheses, citations, or other research-like outputs. This paper argues that the more durable research-system problem is integration. As AI-assisted systems increase the supply of candidate research artifacts, communities must decide which outputs deserve attention, whether their sources and structures are sound, how they relate to existing knowledge, who is responsible for them, whether they can be repaired or should be rejected, and when they deserve downstream reliance. The paper defines this constraint as the integration bottleneck. It develops an artifact-state vocabulary distinguishing generated outputs, candidate artifacts, reviewable artifacts, checked artifacts, integrated artifacts, authority candidates, authoritative objects, rejected objects, and unresolved objects. Three evidence clusters are examined: • Hallucinated citations entering scholarly records. • AI-mediated pressure on peer review and evaluation. • AI-assisted mathematics as a leading case in which generated or AI-associated artifacts move through expert digestion, formal verification, refinement, and authority uncertainty. The argument is methodological rather than alarmist. It does not claim that peer review is obsolete, that AI-generated research is generally invalid, or that all fields face the same burden. Instead, it argues that generation and integration are different research-system functions. The Reflexive Laboratory contributes a vocabulary and architecture for this problem, including: • Transcript-sufficient trace • Governed candidate objects • Source-status discipline • Artifact-integrity checks • Sanity-check operators • Review routing • Authority-state separation The paper concludes that the next phase of AI-assisted research governance should focus not only on what systems can generate, but on how candidate artifacts become usable knowledge. Keywords: AI-assisted research; integration bottleneck; candidate artifacts; scholarly communication; peer review; hallucinated citations; artifact integrity; source verification; proof digestion; formal verification; Reflexive Laboratory