Observed Self-Description Drift in a Vision-Language Model Under Screenshot Analysis

This artifact documents a short interaction sequence with a locally hosted vision-language model (Qwen3-VL-4B) in which the model simultaneously asserts an inability to “see” or interpret images while accurately extracting and summarizing the contents of uploaded screenshots. The uploaded material consists of three screenshots captured sequentially during a single session. In the first, the model provides a capability disclaimer limiting its visual abilities. In subsequent turns, after screenshots are uploaded, the model successfully interprets layout, headings, and semantic content from the images without reconciling the contradiction in its self-description. This artifact is presented as a qualitative observational record illustrating self-description inertia under alignment constraints, where capability execution and capability disclosure diverge. No claims are made regarding general model performance or safety; the screenshots are preserved solely to document this behavior for future analysis.