[qayxc]

This doesn't necessarily relate to the inference itself. No models are exposed to input directly when using web-based APIs, there's pre-processing layers involved that do undocumented stuff in opaque ways.