Gemini 4 Argon fixes off-screen overlay drift in manual test

The test is a direct sequel: the developer had previously run the same screenshots and prompts through Gemini 3.5 Flash Lite and watched a red box meant for the “新規追加” button land outside the screen. That failure pointed to weak pixel-level boundary detection, which made automated manual generation unreliable. This time, Gemini 4 Argon not only placed the boxes correctly but also invented a visual grammar the prompt never asked for — red solid lines for action points, yellow dashed lines for check points — suggesting the model is making document-level judgments, not just following coordinates. The remaining issues are of a different kind: simplified-Chinese-style glyphs in the PDF, banners drawn over input fields, and clipped text at fixed-width banner edges. According to the developer, these are downstream rendering artifacts that added prompt constraints could likely prevent, which would move the workflow from “fix the boxes by hand” to “tell the model how to draw.”