**Some findings from testing Jev , particularly Choice vs Noul and Choice labels** I've been playing around with Jev

**Some findings from testing Jev , particularly Choice vs Noul and Choice labels** I've been playing around with Jev 1.13.0 quite a bit and found some behavior that I think is worth knowing if you're using its probabilities downstream. The biggest takeaway is very simple: The way you represent a question can substantially change the probability Jev returns, even when the underlying question and option descriptions stay semantically same. I started by repeatedly sending identical questions just to see how stable the outputs were. Jev is stochastic (shocker I know): identical calls don't necessarily return exactly the same probability. However, when I repeated questions many times, the average was generally very remarkably stable with very little standard deviation So there seems to be a useful distinction between: individual-call noise vs stable differences caused by changing the question representation + TL;DR: Jev’s probabilities can be surprisingly sensitive to how you represent the exact same question. Noul and equivalent binary Choice questions can give substantially diff