Jev isn't tricked by ordering, and the others are.

TypeSafe's ground-breaking System One model, Jev, has a lot of copycats. They're calling themselves "decision models," which I think is good, because they're decidedly not behaving like System One models in my testing. Most don't describe their architecture, but universally they seem to be built on top of fast LLMs.

Hmmm, I thought. If that's true, they'll probably exhibit some of the weaknesses of LLMs, like the primacy effect and the recency effect. And they sure do.

I'm hosting the results (and the code for reproduction) in this GitHub repo. Let me know what you think!