The recording reference is how every other document in a title chain points at this one. It is a citation, not a database key, and the difference shows up the moment somebody tries to look one up using your normalised version of it.
What is on the stamp
The clerk’s stamp is applied at the moment of recording and carries more than a number. Typically it holds the county and state, the date and time of recording, a reference — book and page, an instrument number, or both — the recording officer’s title, and often the fees collected and any documentary transfer tax. Several of those fields matter to somebody downstream.
The time is the least obvious and among the most important. Recording statutes decide priority between competing interests, and under the race and race-notice regimes used by most states, priority turns on who recorded first. Two instruments recorded on the same day are separated only by the time on the stamp and the sequence in the recorder’s own numbering. An extraction that stores a date and discards a time has thrown away the tie-breaker.
The fee and tax figures are tempting as a route to sale price, because documentary transfer taxes are usually levied as a rate on consideration. Be careful: the rate is set by statute and varies by state and often by city, some conveyances are exempt and stamped as such, and some deeds recite a nominal consideration with the real price elsewhere. Extract the tax amount as an amount, not as a derived price, and leave the arithmetic to somebody who knows which rate applied on that date in that jurisdiction.
Book and page is not unique
The oldest convention is a physical one. Instruments were transcribed into bound volumes; the citation is the volume and the page it starts on. Different jurisdictions call the volume different things — Deed Book, Official Records Book, Volume, and in Maryland and parts of New York the older Liber and Folio — and some counties, having gone through microfilm, cite a reel and image or frame number instead.
The point that breaks pipelines is that these are separate series. A county may have run a Deed Book series, a Mortgage Book series and a Miscellaneous series at the same time, each with its own volume 1. So “Book 1245, Page 331” identifies a document only in combination with the county, the state and the series. A schema of two integers is not an identifier; it is two integers.
{
"recording_reference_raw": "Recorded 04/09/2019 at 2:17 PM\nOFFICIAL RECORDS BK 1245 PG 331-334\nFRANKLIN COUNTY, OHIO",
"jurisdiction": { "county": "Franklin", "state": "OH" },
"series_as_written": "OFFICIAL RECORDS",
"book": "1245",
"page_start": "331",
"page_end": "334",
"recorded_date": "2019-04-09",
"recorded_time": "14:17"
}
Note the page range. Long instruments span pages, and the citation may be written as a range, as a start page only, or as a start page plus a page count. All three occur, and a schema with a single integer page silently truncates the first.
Instrument numbers and the cutover
As recording offices moved to electronic systems, most adopted a sequential instrument number — also called a document number, file number or reception number — usually with a year prefix, such as 2019-0043221 or 201900043221. It is assigned in order of receipt, so it encodes the same priority information the time stamp does.
Every county made that change on its own date, and around that date the records are mixed. Documents recorded before the cutover have a book and page and no instrument number; documents after have an instrument number and, in many counties, a book and page as well, because the office kept assigning them for continuity. So the correct schema has both references as optional and requires at least one, not a single reference field with a type discriminator that has to be guessed — the general case of designing a schema for variants you have not seen.
The formatting traps are mundane and expensive. Leading zeros are significant: 2019-0043221 stored as the integer 43221 no longer matches the recorder’s search index. Hyphens appear or do not, sometimes within the same county. Prefixes are used for document class in some offices. And the year prefix is a two-digit year often enough that a naive parse of 19-0043221 as a year is wrong in both directions.
There is no national registry of these formats. Every recording office publishes its own conventions and its own search interface, and the only authority for what a reference in a given county looks like is that county’s recorder or clerk. Treat any per-county pattern in your pipeline as configuration to be verified against that office, never as something inferred from a sample of documents.
Keep the raw string
Everything above is an argument for the same design: the raw stamp text is the field of record and the normalised fields are derived from it, stored beside it, and allowed to be null.
There are three reasons this is not merely cautious. First, the stamp text is what a person quotes when they cite the instrument, and reassembling “Official Records Book 1245, Page 331” from parsed fields requires knowing the county’s house style, which you do not. Second, when the parse is wrong, the raw string is the only way to find out; a corrected parser can be re-run over stored raw text and cannot be re-run over discarded raw text. Third, stamps contain things your schema does not have fields for — a clerk’s initials, an auditor’s transfer endorsement, a plat reference — and those turn out to matter about once a year.
Stamps are also physically awkward. They are frequently rotated, because the clerk stamped the page sideways in the margin; they are frequently the darkest thing on a light scan, or the lightest thing on a dark one; and they overlap printed text, which is the problem discussed from the other side in extracting grantor and grantee names. Detect and deskew the stamp region separately from the body text rather than running one recognition pass over the whole page; the general handling is in the OCR pipeline page.
Re-recorded and corrective instruments
The case that most often produces a wrong answer with high confidence is a re-recorded instrument. When a deed is recorded with a defect — a missing notarial seal, a typo in the legal description — it is commonly re-recorded, sometimes with a legend across the first page reading “RE-RECORDED TO CORRECT THE LEGAL DESCRIPTION”. That document now bears two recording stamps, usually months apart, and both are real. Each carries its own date, so both go through the same date-field validation as any other extracted date.
A pipeline that extracts “the” recording reference from such a page will pick one, and which one it picks depends on stamp position rather than on meaning. The correct output is a list of recording events with an order, plus the re-recording legend if present. The earlier stamp is the original recording; the later one is when the corrected instrument entered the record. Which of them controls for any particular purpose is a question for a title examiner, and the extraction’s job is to make sure the examiner can see that there were two.
Related but distinct is a separate corrective deed or scrivener’s affidavit: a new instrument with its own recording reference that recites the reference of the deed it corrects. That recited reference is a fourth kind of string on the page — a citation to another document, not to this one — and merging it into the recording-information field is a common and confusing error. Extract references to other instruments into their own array with a relationship label.
Top comments (0)