Drone-based applications such as navigation, inspection, and search depend on vision-language models (VLMs) that can answer spatial visual questions. The VLM faces three challenging tasks: grounding each referenced phrase to a scene object, choosing a viewpoint for spatial relations, and verifying the relation against accurate 3D geometry. In current frozen VLMs, all three tasks happen inside a single forward pass, and only the final answer is visible from outside. As a result, when that answer is wrong, we cannot tell which task failed. To address this, we introduce AeroVerifier, which exposes each of these three tasks as an explicit and checkable interface. Given a question, AeroVerifier compiles it into an executable typed program, runs the program over object-level 3D evidence, and returns the program answer together with stage-wise diagnostic signals to the base VLM as single-step revision feedback. On a 1000-row Open3D-VQA subset, both AeroVerifier’s program answer and its revision feedback yield higher answer accuracy than the Qwen2.5-VL-7B and GPT-5.4 API baselines.