Primary care
clinical coding.
With evidence.

300 synthetic letters carrying 3,600 planted identifiers were cleaned by the local de-identification step (NHS rules plus a clinical PII model), on 23 October 2026.

Synthetic letters with planted identifiers, not real correspondence. Trust names, clinic dates, age, gender and occupation are deliberately kept. An earlier run found 54 of 1,800 identifiers left readable, all in unusual date and surname formats; the rules were fixed and this run, on a fresh seed, found none. Not a guarantee for real letters.

Cleaning happens on the practice's own machine, before anything is sent.

Identifiers removed — 3,600/3,600: Names, NHS numbers, dates of birth, postcodes, phone numbers and more were planted in 300 synthetic letters. None was left readable in the cleaned text.

Letters fully clean — 300/300: Every one of the 300 letters came back clean, including the patient's name.

Clinical wording intact — 1,861/1,861: Diagnoses, medicines and findings were checked after cleaning. None was damaged.

23 synthetic letters tested the coding workflow. These are the recorded development results from 9 October 2026.

Cases were reused during tuning with an in-house answer key. These expected-item counts are not overall accuracy or independent validation. Expected uncoded handling was 31/46.

A suggestion still needs clinical review.

Clinical fact recall — 153/153: Did the expected item appear in the extracted results?

SNOMED concept matching — 106/107: One incorrect concept remained. The anaphylaxis example lost detail in its suggested wasp-sting concept.

Clinical context — 130/132: Active, historical, suspected and negated meanings were checked. Two assertions were incorrect.

The local PP-OCRv6 recognizer was tested on selected lines from 10 public de-identified letter documents new to that recognizer benchmark, on 1 October 2026. The test used very low-resolution, small-size WebP files.

Manually selected line crops, not automated whole-page recognition. An 11th new photographic example could not be localized for this line test. Earlier project use or pretraining overlap cannot be excluded. Related and previous examples are counted separately.

Reading text correctly is one part of coding.

Exact-line recognition — 70/77: Recognition was compared with a frozen visual transcription of the selected clinical lines.

Clinical field retention — 74/77: Clinical probes included names, doses, units and context wording. Exact-line and field checks ask different questions.

Negation recognition — 16/18: Two negation probes were missed. These OCR checks do not establish SNOMED mapping, medication association or clinical safety.