Designing the content component of immersive blended science learning: a CAMIL-grounded framework refined and pre-validated by a simulated multi-model Delphi panel

Immersive technologies - virtual reality (VR), augmented reality (AR) and 360◦ video - are spreading through school science, yet teachers lack research-informed criteria for deciding what content to deliver immersively and how to embed it in blended lessons. Building on a prior conceptual analysis, this paper develops the content component of an immersive-learning methodology into an explicit design framework grounded in the Cognitive Affective Model of Immersive Learning (CAMIL). The framework was refined and stress-tested with a novel instrument: a simulated Delphi panel of fourteen large language models (LLMs) from eight model families and three providers, each adopting an expert persona, across three rounds (generative development, rating, and re-rating after anonymised feedback). The panel expanded the framework from eleven to fourteen content-design criteria and a VR/AR/360◦ modality-fit mapping. We report the exercise transparently, including its limits. Relevance ratings sat at a ceiling - every item was endorsed - so the content-validity indices are non-discriminating, and the panel, which also generated the items it rated, offers refinement and pre-validation, not independent validation. The informative signal lay in feasibility, where the panel discriminated sharply and, on re-rating, became markedly more pessimistic: full VR, differentiation and teacher orchestration were judged least feasible in real classrooms. We contribute (i) the refined framework and (ii) the simulated multi-model Delphi as a fast, fully logged but explicitly synthetic pre-validation method, whose affordances and limits for educational design research we analyse. No human participants or classroom data are involved; the LLM panel complements rather than replaces human expertise.