Technology · Correction design
The real product begins
after the misread.
Glare, plate frames, stacked characters, dirt, motion, temporary tags, and unfamiliar formats guarantee disagreement. A plate-driven operation is therefore a correction system with recognition at the front.
The test
A wrong answer should remain cheap to fix.
The dangerous failure is not a reader returning nothing. It is a plausible wrong plate becoming a durable identity before the person holding the camera has a natural chance to catch it.
Do not reduce this question to an accuracy percentage. Field conditions, plate populations, device models, and acceptance thresholds change what any single number means.
The correction loop
Uncertainty should stay visible until it is resolved
Prefer no answer to false certainty
Confident-looking guess
RISK- Looks settled before a human checks it
- Can match the wrong vehicle or resident
- Makes downstream correction expensive
Visible uncertainty
RECOVERABLE- Invites another frame or a quick edit
- Keeps the operator in the same workflow
- Separates candidate from accepted identity
Correct in context
→A correction is not merely changing text. Any decision that used the old plate must be reconsidered.
Measure the correction system
A recognition team should study what operators had to repair, not only what a labeled test set says.
Rate
Manual correction
How often did a valet change the candidate before acceptance, by device and operating condition?
Shape
Character disagreements
Which positions, formats, frames, glare patterns, or temporary plates create recurring trouble?
Cost
Time to recover
How many actions and seconds did correction add to the arrival, and did it force a restart?
Escape
Errors found later
Track plates corrected after check-in separately. Those are the failures the first-line workflow did not catch.
→Publish methodology before publishing a headline accuracy number. In this article, we publish neither Valletto's nor anyone else's percentage.
The product test
Deliberately feed it the hard plate.
Use glare, an angled camera, a frame that obscures a character, and a plate format outside the local norm. Watch whether the system communicates uncertainty, preserves typing, and re-evaluates downstream matches after correction.
The takeaway: the best recognition system is the one that makes its inevitable mistakes obvious, reversible, and useful for the next release.
