Set the review gate. Watch what still reaches a user.
a RavnLab demonstration · pairs with the Confident-Wrong Checklist and the Plausible-Wrong Benchmark
how to read it
Each dot is one answer the system produced
Left to right is how confident the model was. Blue dots were right; red dots were wrong. The gate is the vertical line: everything to its right gets shipped automatically, everything to its left goes to a human to check. You are tuning one number - where to put the gate - and trading two costs against each other: wrong answers that reach a user, and the human effort of review.
the simulation
Move the gate
Its confidence tracks its accuracy: when it says 90%, it is right about 90% of the time. Most wrong answers sit on the left, at low confidence.
Illustrative data - the shape of the problem, not one model's scores. A demonstration we produced, no client involved. The pattern is the human-review gate from our reliable-agent-patterns work.
what it shows
You cannot threshold your way out of miscalibration
With the calibrated model, the confidence number is a real dial: raise the gate and you reliably ship fewer wrong answers, because the model is rarely confident about a wrong one. With the overconfident model, the wrong answers hide among the confident ones - at the same gate it ships far more of them, and no setting sorts them out unless you send almost everything to a human. The gate is only as good as the confidence behind it. That is what calibration buys, and it is measurable.
- Confidence is only useful when it is earned. A number that does not track accuracy is worse than none, because it invites exactly this false sense of safety.
- Every red dot on the right is a confident-wrong answer. The six ways they get produced are in the checklist, each with a test you can run.
- Measuring calibration on adversarial, sourced cases - the overconfidence gap, refusal correctness - is what the public benchmark does.