English
Official papers, read field by field
A certificate or a permit is not prose: it is a set of labelled fields, some printed, some written in, some stamped across at an angle by a third office months later. Getting it into text means keeping each value with the label it belongs to — surname, number, date of issue, place — and not letting a rubber stamp bleed into the line underneath. Because these papers carry personal data, the samples on this page are historical documents of people long dead, or blank and specimen forms.
Sample pages
Real pages from public collections, shown beside their transcriptions. Pages whose transcription is still being checked are marked.
Source · Public domain · British Mandate authorities; the holder died long ago
Transcription being verified
A single archival document with no published transcription. It is included because four different offices wrote on it, in three languages, over a period of weeks.
What the model is told to watch for: Treat names, document numbers, dates, coded fields, nationality abbreviations, stamps, signatures, and machine-readable zones as separate structured elements. Distinguish printed data from handwritten amendments and ignore guilloche patterns, holograms, watermarks, or security microprinting unless they contain legible textual information.
United States Passport Office records, 25 May 1905 · Source · Public domain
Transcription being verified
Indexed by name and date in the National Archives passport application series, but not transcribed in full. Colnik died in 1958.
What the model is told to watch for: Treat names, document numbers, dates, coded fields, nationality abbreviations, stamps, signatures, and machine-readable zones as separate structured elements. Distinguish printed data from handwritten amendments and ignore guilloche patterns, holograms, watermarks, or security microprinting unless they contain legible textual information.
US Federal naturalization records, California, 5 February 1904 · Source · Public domain
Transcription being verified
Federal naturalisation records are indexed by name but not published as text. The value of this page is the alternation of printed clause and handwritten answer within one sentence.
What the model is told to watch for: Treat names, document numbers, dates, coded fields, nationality abbreviations, stamps, signatures, and machine-readable zones as separate structured elements. Distinguish printed data from handwritten amendments and ignore guilloche patterns, holograms, watermarks, or security microprinting unless they contain legible textual information.
Wellcome Collection, c. 1918 · Source · CC BY 4.0 · Wellcome Collection, London
Transcription being verified
The instructions are printed matter reproduced in millions of copies; they are not available as text. The image has been cropped so that the named holder’s details do not appear.
What the model is told to watch for: Treat names, document numbers, dates, coded fields, nationality abbreviations, stamps, signatures, and machine-readable zones as separate structured elements. Distinguish printed data from handwritten amendments and ignore guilloche patterns, holograms, watermarks, or security microprinting unless they contain legible textual information.
How an official document is put together
The family is wide: birth, marriage and death certificates, naturalisation and citizenship papers, passports and travel permits, licences, registration cards, ration and identity books, service records, court and notarial instruments. Nearly all of them are a printed frame with a small amount of variable data dropped into it. The frame carries the issuing authority, the form number, the legal wording and the field labels; the data is typed, written or later machine-printed on the rules between them. A serial number is usually printed in a different colour or by a different press, so it is often the sharpest thing on the page and the easiest to misread by one digit.
On top of the frame sits everything added afterwards. Rubber stamps record that the document was seen, entered, filed or paid for, each with a date in its own format and often across text that was already there. Embossed or wet seals raise the paper and disappear in flat light. Endorsements, visas and countersignatures crowd the margins, sometimes in another language and another script. Older papers may have printed clauses struck out where they did not apply, and amendments written in and initialled — those deletions are part of the meaning, not damage.
Modern identity documents add a machine layer. A passport data page carries two lines of the machine-readable zone in OCR-B, in which the country and nationality are three-letter codes, dates are six digits, the name field uses double chevrons between surname and given names, and filler chevrons run to the end of the line; the last character of each group is a check digit. Around it are guilloche patterns, microprinting, holograms and a ghost portrait, all designed to defeat copying — and all of which the reading has to ignore rather than turn into characters.
Why ordinary OCR struggles here
Security printing is designed to interfere
Guilloche backgrounds, rainbow tints, watermarks, microtext and holographic overlays exist precisely to make reproduction hard. They sit under the data, so a reading has to separate the value from the pattern, and holograms on a card change what is legible with the smallest change of angle.
Fields, not sentences
Surname, given names, place of birth, date of issue, authority: these mean nothing without their labels, and the labels may be above, beside or a long way from the value across a dotted rule. A transcription that returns a list of words in reading order loses which value belongs to which field.
Stamps, seals and endorsements over text
A violet stamp landing across a handwritten line obscures part of both. The date inside the stamp matters, the text underneath matters, and the two have to be reported separately rather than merged into one confused string of characters.
Dates and codes that look like ordinary numbers
A date may be written 05/06/23, 5 JUN 23 or 230605, and only the document tells you which order the parts run in. Nationality codes, form numbers, batch marks and check digits are all bare alphanumerics, and a single wrong character in a document number is worse than an obvious gap.
Who works with this material
Archives and record offices
Registers, licences and permits are digitised by the metre, and item-level indexes are what make them findable. Pulling names, dates and reference numbers out as fields turns a shelf of images into a catalogue that a reader can search.
Family historians
Certificates and immigration papers are the load-bearing evidence in a pedigree. Exact spellings, ages, occupations and place names matter, and a note of what is illegible is more useful than a confident reading that quietly settles the wrong ancestor on a family.
Legal and immigration case workers
A case file arrives as photographs of documents in several languages and formats. Getting a consistent set of fields out of them, with stamps and endorsements recorded separately, is what makes a chronology possible.
Back-office teams handling verification
Onboarding and compliance work involves reading data pages and certificates at volume. Here the requirement is as much about handling as about accuracy: only the fields that are needed, held only as long as necessary, with the rest never leaving the image.
Getting the best transcription
Decide what you are allowed to keep
Before anything is uploaded, settle which fields you actually need and which should be masked or dropped. It is usually possible to work from a document number and an expiry date without ever extracting a full name, an address or a photograph.
Photograph flat, in even light, both sides
Lay the document on a plain surface, avoid direct light on laminate or a hologram, and capture the back as well — endorsements, continuation stamps and the machine-readable zone often live there. For a booklet, photograph the opening flat rather than at an angle.
Name the fields you want returned
Ask for a field list — surname, given names, document number, date of issue, expiry, issuing authority — rather than a block of text. Say which date format you want, and ask for the value as printed alongside any normalised form so the reading can be checked.
Treat numbers as needing a second look
Document numbers, reference codes and check digits are exactly where a single-character error does the most damage. Have them read back a second time, compare the two readings, and flag any disagreement instead of accepting the first result.
Frequently asked questions
- Can it read the machine-readable zone on a passport?
- The two lines of OCR-B can be transcribed, including the chevron fillers, the three-letter country and nationality codes and the six-digit dates. Because each group ends in a check digit, the values can be checked arithmetically, which is the quickest way to catch a misread character.
- How is personal data handled?
- Extract only the fields you need and mask the rest — cover a photograph, a signature or an address before uploading, or ask for those fields to be omitted from the output. Every sample shown on this page is a historical document whose subject died long ago, or a blank or specimen form.
- Are stamps and seals transcribed?
- Yes, as separate elements. A rubber stamp is reported with its text and any date, an embossed seal is described where it cannot be read, and both are kept apart from the text they were pressed across so that neither corrupts the other.
- What about handwritten amendments to a printed form?
- They are marked as handwritten and kept distinct from the printed frame. A struck-out clause is reported as deleted rather than dropped, which matters, because on many older forms the deletions are how the document says what it means.
- Can it work from a photograph taken with a phone?
- Usually, provided the whole document is in frame, the lighting is even and the plane of the page is roughly parallel to the camera. The common failures are a flash reflection across a laminated card and a corner cropped off, both of which are cheaper to fix by retaking the picture.
- What if the document is in two languages or two scripts?
- Bilingual certificates and permits are common, and each version can be returned as its own block with its language named. Non-Latin script on the same page — an Arabic or Hebrew endorsement beside English — is transcribed in its own right rather than transliterated by default.
Related scripts and pages
Read official papers as data
Upload a certificate, permit or archival record, choose Official and Identity Documents, and name the fields you want — and the ones you would rather never leave the image.