i kept returning to the question of what counts as a verified output when a single inference request returns more than text.

OpenGradient’s SDK supports native image-output models. Generated images are returned separately through "result.images" as data URIs, while any accompanying text appears in the normal response.

The documentation also makes an important boundary explicit: those images are delivered out of band and are not included in the signed output hash.

At first, that sounded like a minor implementation detail.

It isn’t.

The signed response can provide cryptographic evidence for the output covered by that hash, but the same signature does not independently authenticate the generated image bytes.

That does not mean the image was altered or delivered incorrectly. It means the image and the accompanying text do not receive equal verification coverage from the signed-output mechanism.

That is the distinction I keep coming back to.

Users experience one multimodal answer, but the SDK delivers separate artifacts with different evidence boundaries.

This could matter when the visual itself carries the instruction, decision, or claim that later needs to be audited. An authenticated text response cannot, by itself, prove that the displayed image is the exact image produced during the same inference.

Does out-of-band image delivery keep multimodal inference practical, or leave an important gap in what applications can cryptographically verify?

OpenGradient can return generated images separately from the signed text output, but those image bytes are not included in the signed output hash.

Does this keep multimodal inference practical, or create a real verification gap?

#OPG @OpenGradient $OPG $TNSR $UB