The setup

One document type, nine fields, three of them handwritten: the owner's phone, their email, and the pet's weight. The destination is a spreadsheet. The specification says what each column is, which ones are handwritten, and the bounds a sensible value sits inside. Before any reading, the correct values for all thirty forms were typed by hand into a truth file, one row per document. That file is what makes an accuracy number possible at all.

Two rules from the handwritten-cards job are built in. The reader is told to copy what is written and never to fill anything in, so a date with no year stays a date with no year and is flagged, instead of a confidently wrong one. And the reader is never asked how sure it is, because on that job every wrong amount arrived at high confidence. Each form is read three times, and a value goes in untouched only when all three reads agree and it passes its column's checks.

The accuracy check, thirty documents

FieldRightWent straight inSilent errorsOne read aloneTier
Owner name30 of 3093%01trust
Phone (handwritten)28 of 3080%03glance
Email (handwritten)30 of 3090%01trust
Pet name30 of 30100%00trust
Species30 of 30100%00trust
Date28 of 3093%00trust
Weight (handwritten)27 of 3063%17check
Vaccines up to date30 of 30100%00trust
Reason for visit30 of 30100%00trust

"Went straight in" is the share of values that needed no person. "Silent errors" is the number that went in wrong without a flag, the only column that matters. "One read alone" is what the silent-error count would have been with a single read and no agreement rule.

With one read alone, twelve wrong values go in silently. Requiring agreement brings it to one. Twenty-four cells across eighteen documents are flagged for a person, which is the price of that.

The one it could not catch, on purpose

Agreement cannot catch a misread that happens the same way every time. The demo plants exactly that: a handwritten 7 that every read sees as a 1. All three reads agree, the value passes its bounds, and it goes straight in. Only the accuracy check against the hand-typed truth finds it, which is why that check is not optional and why every column gets a tier from it.

  • Trust: no silent errors in the sample. The pipeline fills it; nobody looks.
  • Glance: no silent errors, but enough flags that a quick look at the flagged ones is worth it.
  • Check: even one silent error in the sample, so a person checks every value in that column. Here, the handwritten weight.

That last tier is handed over as a recommendation, in writing, with the number beside it. A tool that said "90% accurate" and stopped would have hidden the one column where 90% is not good enough.

Why this matters to you

Every tool in this category demos the happy path. The useful question is not whether software can read your paperwork; it is how often it reads yours correctly and what happens the rest of the time. The Capture Sprint answers that on your own documents, with the table above filled in for your fields, before you decide how much to rely on it. You can run the method above yourself in about a second; the sprint is the version on your paper, into your system.

Tell me what gets retyped