Every cost tool tells you what you already spent. This one prices the call first, states a ceiling it can be held to, and decides whether to allow it. Predict below, then fire the real call and see how close it was.
Every real call run here is recorded and checked against its own ceiling. This is the running score — not a claim, a tally.
The bound is calibrated by conformal prediction, so it needs history before it will promise anything — below nine real calls it says so rather than inventing a number. It tightens as data arrives without giving up coverage.
I'm researching how teams handle agent spend — particularly context that gets re-billed every turn. If that's a problem you recognise, I'd value 20 minutes. No pitch.
Prefer to talk in the open? Join the Discord to share feedback.