Expert teachers possess rich tacit knowledge for evaluating preservice teachers' classroom performance — knowledge that is highly intuitive, context-dependent, and difficult to articulate. This challenge is particularly evident in complex domains such as classroom management, where assessment criteria must often be constructed rather than retrieved. Although generative AI (GenAI) is increasingly used in educational assessment, it remains limited in accessing such embedded expertise. A critical question therefore arises: how can expert tacit knowledge be systematically externalized in ways that enhance both AI capability and expert professional learning (Nonaka & Takeuchi, 1995)? This study proposes CADET (Cognitive Apprenticeship-Driven Expert–AI Dual Training), a structured dialogue framework grounded in cognitive apprenticeship (Collins et al., 1991). CADET conceptualizes GenAI as a novice learner that acquires expert-level assessment knowledge through progressive, dialogue-driven interaction. The framework is being prepared for implementation in a VR-based classroom management context, where expert teachers evaluate preservice teachers' responses to simulated disruptive student behaviors. CADET operates through three progressive phases: Modeling–Acquiring, Coaching–Enacting, and Monitoring–Applying. Central to the framework is a novel double-transfer mechanism. First, experts externalize tacit evaluative reasoning, enabling GenAI to internalize expert-like decision-making. Second, GenAI functions as a cognitive mirror, prompting experts to examine and refine their implicit reasoning through structured questioning — a process termed reverse articulation (Schön, 1983). This reciprocal process reframes expert–AI interaction as a site of mutual learning. CADET positions GenAI not merely as an assessment tool, but as an active participant in knowledge co-construction, while simultaneously supporting experts in making tacit reasoning more explicit. Empirical validation is planned to examine the conditions under which dual knowledge transfer can be effectively achieved.