TasksDocument ExtractionBusiness track

Intake form with handwriting-style noise

biz-extract-02suite v1.0json

Quality15%judge panel
Correctness70%deterministic · field-match
Completeness15%deterministic · schema-validate

Exact prompt sent to every model

Below is an OCR transcription of a handwritten new-patient intake form, including OCR errors and crossed-out corrections marked with strikethrough. Extract to JSON with fields: full_name, date_of_birth, phone_primary (E.164), phone_alt, email, address {street, city, state, zip}, insurance {carrier, member_id, group_number}, emergency_contact {name, relationship, phone}, allergies (array of strings, empty if "none"), medications (array of {name, dose}), consent_signed (boolean), signature_date. Where a value was corrected, use the correction. Where OCR clearly garbled a value, repair it only if the intended value is unambiguous, otherwise null. Return only the JSON.

A private fixture (biz-extract-02/intake.txt) is shown to the model but not published.

Responses

Latest published run: gemini-3.8-flash. For each model the repeat closest to its published median is shown; every repeat is in the repository.

No published attempts for this task yet

When the run is published, every model's raw response, per-dimension scores, judge rationales, human review notes and rendered artifacts appear here side by side.