I turned the list scanner on
I left it off in production until I was sure. This week it learned from corrections without knowing whose they were, and I switched it on.
When I finished building the scanner for handwritten lists, I decided to leave it switched off in production. It worked and it was measured, but “works in my tests” isn’t the same as “works with other people’s handwriting.” I preferred to turn it on once I had a reason to trust it.
That reason came from a direction I didn’t expect: corrections. When someone reads a list and changes a misunderstood word, that correction teaches something. I made the scanner learn from them, but without knowing who made them: what gets stored is the misread word and its fix, not the person or their list. Over time, the scanner makes fewer mistakes with the abbreviations and everyday expressions people actually write in Peruvian lists.
I also gave it a behavior that matters to me: it can suggest a word, but never silently. It always shows you what it understood before adding it.
On October 11 I turned it on in production. There’s still a monthly cap on photos, because Nada Menos is free and every photo costs something, and the switch to turn it off is still there in case it’s needed. Now tell me how it does with your handwriting.