The Verification Depth Gradient: A Five-Stage Model of Citation Practice in the Age of AI

This paper proposes a five-stage Verification Depth Gradient that reframes the "vibe citing" problem-large language models' tendency to generate plausible but fabricated citations (Adams, 2026)from a technical error classification to a behavioural model of human citation practice. The gradient ranges from deep primary-source engagement (Stage 1) through AI-assisted retrieval with human verification (Stage 3) to fully automated citation without human oversight (Stage 5). The model's central claim is that the critical transition-the point at which citation contamination risk shifts from negligible to structural-lies not between human and AI citation, but between Stage 3 (where human anomaly-detection capacity remains active) and Stage 4 (where it ceases to function).

This transition is externally invisible: no feature of a submitted manuscript reveals the author's verification depth. The paper situates this invisibility within the HYC Theorem framework (Yamada, 2026d) and discusses the NeurIPS 2025 editorial response as a real-time institutional application. Implications for bibliometric measurement infrastructure and editorial policy are discussed.