Jev: new paradigm, or just logits with a better marketing team?

Jev is being pitched as a brand new kind of model: give it a fixed set of choices, get calibrated probabilities back in one shot.

but isn't that just a forward pass and a softmax? you can already do that fully on-device:

• NobodyWho: open-source, drop-in local LLMs for basically any stack (Python, Swift, Kotlin, Flutter, RN, Godot)
• Cactus: on-device inference for phones, already routes by model confidence
• llama.cpp: the OG, grab the logits yourself

I get that the secret sauce is supposed to be calibration, and sure, raw logits from a small model aren't calibrated and usually have a favorite letter. but is that a real moat?

what am I missing? has anyone actually compared them?