用思维反哺机制让AI帮人理清研究逻辑，防止盲目依赖。

Used as a capable servant, generative AI has greatly accelerated intellectual work, yet it also risks eroding human epistemic agency by encouraging uncritical acceptance of AI-generated reasoning. Preserving that agency calls for mechanisms that augment human metacognition during AI-assisted work. We therefore propose the Synthesis-Analysis Reciprocity Model, which views intellectual construction as a reciprocal interaction between two cognitive functions. Synthesis selects and combines the components of the artifact; Analysis evaluates them critically against objective indicators and constrains the Synthesis that follows. Grounded in this model, we present the Vibe Compiler, a research-logic compiler that helps researchers turn vague intuitions (Vibes) into coherent research logic. The system attempts to compile those intuitions against a paper ontology of 16 academic parameters. It treats compilation failures as signs that logical components are missing. Rather than fill those gaps autonomously, it returns reflective questions that prompt researchers to develop the missing reasoning themselves. We further characterize the origins of structural gaps along two orthogonal dimensions: cognitive function (Synthesis versus Analysis) and executing agent (human versus AI). The four resulting types of origin give a principled way to identify where breakdowns in intellectual construction arise. Crucially, our design implements the type in which the AI probes its own synthesized output, itself driven by the user's Vibes, and thereby stimulates human metacognition. This choice raises researchers from passive "Makers" of the output into "Managers" who critically direct and validate what the AI produces. In a prototype on NotebookLM and Gemini, AI behavior depended less on prompting than on the structure of the knowledge supplied. The framework spans a learner layer and a researcher layer.