Thesis
The interface can make model-produced grounding inspectable, while the coordinate convention remains an application-level assumption rather than a documented model contract.
Observed facts
The pull request describes a Bun app that overlays structured boxes and exposes label, confidence, and short reasoning when a box is selected.
1Replay mode is the default, requires no API key, and uses synthetic answers labeled illustrative; live mode supports Vercel AI Gateway, Mistral API, and OpenRouter.
1
Mechanism
Because the author reports no box convention in Mistral’s launch materials or model card, the prompt requests normalized 0–1000 coordinates and the app parses and renders them.
Why now
The demo turns an underspecified model response into an inspectable interface, making the gap between a compelling visual result and a stable integration contract visible.
Implications
Teams can prototype grounding interactions, but should validate coordinate order, origin, scaling, and behavior near image boundaries before relying on overlays.
Evidence boundary
The source documents a community demo and its stated implementation; it does not establish an official schema, general model accuracy, or production reliability.
Watch conditions
Look for an official coordinate specification or independent tests across varied images and boundary cases.