Demo 1 of 2
Set the promise
Pretend you run the toll road. Decide how many wrong bills you are willing to accept, and see what TULA does with real cars from fog, clear weather and Indian plates. Every number is recalculated in your browser when you move the slider.
What you are looking at
Its position is the car's doubt: far left, everything agrees; far right, nothing does. Green dots would be billed correctly, red dots would bill the wrong person.
Using other cars whose right answer is known, TULA draws a line at the highest doubt that still keeps your promise. Cars left of the line are charged.
The dots are cars TULA had not seen when it drew the line, so the numbers show how the promise holds up on new traffic.
For the curious: the exact rule and what "on average" means
"On average" means over many calibration sets: one split on its own can land a little above the promise, and the average over splits stays under it. "90% confidence" uses a Clopper-Pearson bound instead, so each split stays under the promise with probability at least 90%, at the cost of charging less.
Now compare
Why not just one "90% sure" rule?
This is how most billing works today: charge whenever the model says it is at least this sure. Move the slider. Orange is that one rule, used for every photo. Green is TULA at your promise above. The red mark is the promise: past it, more innocent drivers are billed than promised. Notice that the same setting is safe in fog and unsafe in clear light.