and finds out if executing a prompt constitutes security risk. surprised that noone attempted something like that already

but, honestly, i think that simple models are to simple to undersrand which prompt goes where, and if you're using expensive model for that, it makes no financial sense

decision models may work well there (The Jevs)

https://sakana.ai/fugu-release/

There’s like fifteen companies offering this…

https://github.com/NVIDIA-NeMo/Switchyard

Leaderboards shift as new models ship, and a February 2026 analysis found high saturation across nearly half the benchmarks reviewed. Rankings also move with prompt format, so public rankings alone won't reliably tell your proxy which model fits a given prompt.

What is Jev

yes certainly - LLM offline in andriod doing TTS STT translation already

Jev

the hard part isn't routing, it's knowing when the router failed: a complex task silently shipped to a cheap model produces an answer that looks normal and isn't. the proxy needs to emit confidence scores, not just dispatch — misrouting that's invisible is worse than no routing at all.