Same AI. Same file. Different device. Different answer.

We ran the exact same medical AI model file (BioMistral-7B, 4-bit) on two machines: a laptop and an Android phone.

Same model. Same questions (200 US medical licensing exam questions). Same settings.

Exam score: 52% on the laptop, 50.5% on the phone. Basically identical.

But 1 in 20 answers changed.

One of them:
A newborn is shaking, sweating, crying high-pitched and won't feed. His mother took codeine syrup throughout pregnancy.

💻 Laptop: treat the opioid withdrawal ✅
📱 Phone: start IV antibiotics for an infection ❌

Nobody changed the model. Nobody retrained it. It just ran on different hardware, and the tiny differences in how each chip does maths added up to a different decision.

This is why "we tested the model" isn't enough. You also have to test it where it actually runs: every device, every update, every speed-up.

That's what Eppy does: it finds the fastest way to run a model on each device, then checks which answers changed, not just the score.

👉 Test the AI where it runs, not just where you built it.

Research benchmark on exam questions, not medical advice.

Changing your AI? We’ll find what broke. myturn@ephard.ai