Can AI Read a Handwritten Medical Record? We Tested It — Including the Page It Couldn't Read.

Guidon Insights · September 2026 · Test pages published below. All content synthetic; no real veteran's records are ever used for testing.

A state veterans agency asked us a question every office with older files eventually asks: "Can it read handwritten documents?" A lot of the records that decide claims — service treatment records from the 70s and 80s, clinic notes, buddy statements — are handwritten, and a tool that only reads typed text solves half the problem. We gave them the honest answer on the call, and then we did what we always do with an honest answer: we built the test that would prove or embarrass it.

The test

We generated three synthetic handwritten pages — a print-style clinic progress note, a cursive nursing note, and a deliberately ruined page: faint, water-stained, scrawled. Then we ran them through the same document reader our pilot offices use in production, as single photographed pages and inside a multi-page scanned packet.

Synthetic handwritten progress note in legible print-style handwriting — **Page 1:** print-style handwriting — a synthetic clinic progress note.

Synthetic nursing triage note in cursive handwriting — **Page 2:** cursive — a synthetic nursing triage note.

Deliberately illegible water-stained page with faint scrawl — **Page 3:** deliberately illegible — faint ink, water stains, scrawl.

The results

Page	What happened
Print-style handwriting	Fully read. All 6 of 6 ground-truth clinical phrases recovered — conditions, exam findings, the plan, even "McMurray equivocal."
Cursive	Read. 3 of 3 ground-truth phrases recovered. Quality varies with the writing — but ordinary cursive held up better than we promised.
Illegible page	Refused — loudly. 13 stray characters recovered, correctly treated as unreadable. And this is the result we care most about:

When a page can't be read reliably, the analysis doesn't skip it quietly. It writes this into the case, where the officer will see it:

[OCR NOTE: page 3 of this document was not reliably readable — it may be blank, or too blurry, faint, or damaged to read. Its contents are NOT included in this analysis.]

A missing page named is a page the officer can chase — re-photograph it, request a better copy, ask the veteran. A missing page silently skipped is a condition that never gets claimed, and nobody ever knows why.

The bug the test caught

Here's the part worth being honest about: when we first built this test, the system failed it. An unreadable page in the middle of a scanned packet was being dropped from the analysis without a trace — exactly the silent skip we tell offices we never do. And a page that yielded a few characters of garbage slipped past the first fix by technically not being empty. The test caught both. The fixes shipped the same day, and the fail-loudly behavior above is now proven by a repeatable test, not asserted by a brochure.

The honest limits: reading quality varies with the handwriting — a doctor's genuine scrawl from 1978 may land anywhere between our page 1 and page 3, and the only real answer for an office's own files is to test on their kind of material, which is exactly what a pilot is for. Scanned packets are read at up to 300 pages per document in the current build, and when a document exceeds that, the report says so explicitly. Synthetic handwriting is cleaner than real handwriting; treat these results as the mechanism working, not a guarantee for every page ever written.

The standard isn't "reads everything." Nothing reads everything. The standard is: reads what's legible, names what isn't, and never silently guesses.

This piece pairs with our 6,000-page capacity benchmark — together they answer the two questions every office with big, old files asks first. Want the test run on your kind of records — redacted or synthetic, your formats, your handwriting? That's a standing offer in every free pilot.