Abstract
Scientific AI can retrieve, synthesize, and reuse claims far faster than any researcher can inspect them. The problem is that provenance is often preserved more faithfully than evidential strength. Weak, biased, or fabricated findings may therefore move through training corpora, retrieval pipelines, and knowledge graphs without the methodological context that should constrain their interpretation. We argue that machine-readable knowledge should be built around layered evidence records rather than papers or extracted sentences alone. These records should retain source text, study design, population, controls, uncertainty, causal status, independent validation, and correction history. The central challenge is not only whether AI can read more literature, but whether it can retain why one claim deserves more confidence than another.