Grantor and grantee are not two name fields. They are one relationship with a direction, and the direction is what the county recorder indexes on. Get it backwards and every chain of title built on your data runs the wrong way.
Direction is the whole field
The grantor conveys; the grantee receives. Recorders have maintained two indexes on that distinction for as long as there have been recording offices: a grantor index, usually called the direct index, and a grantee index, usually called the indirect or reverse index. Title work runs backwards through the grantee index — find the instrument in which the current owner was the grantee, note who the grantor was, then find the instrument in which that party was the grantee, and so on back through the chain.
This is why a swapped pair is worse than a missing pair. A missing pair is a gap somebody notices. A swapped pair produces a chain that links up cleanly and describes a sequence of conveyances that never happened. Nothing downstream can detect it, because both names are real and both appear on the document.
The same relationship appears under other names on other instruments and the roles do not map one-to-one. A mortgage has a mortgagor (the borrower, who grants the security interest) and a mortgagee. A deed of trust has three parties: a trustor or grantor, a trustee who holds bare legal title, and a beneficiary who is the lender. If your schema has only grantor and grantee, a deed of trust does not fit in it, and forcing the beneficiary into the grantee slot loses the trustee entirely. Type the instrument first, then extract the party roles that instrument type actually has.
Anchor on the granting clause
Deeds are formulaic, and the formula is what to anchor on. The granting clause is a single sentence built around an operative verb phrase, and the parties sit on either side of it:
THIS INDENTURE, made this 3rd day of April, 2019, between
ALEX RIVERA and JORDAN RIVERA, husband and wife, of Franklin County,
party of the first part, and
MERIDIAN HOLDINGS LLC, a Delaware limited liability company,
party of the second part;
WITNESSETH, that the said party of the first part, for and in
consideration of Ten Dollars ($10.00) and other good and valuable
consideration, does hereby GRANT, BARGAIN, SELL AND CONVEY unto the
said party of the second part, its successors and assigns forever,
the following described real estate ...
The party before the verb grants; the party after “unto” receives. That holds across the vocabulary variants: a warranty deed says “grant, bargain, sell and convey”, a quitclaim says “remise, release and forever quitclaim”, a special warranty deed narrows the covenant but keeps the same shape. Older instruments use “party of the first part” and “party of the second part” as a layer of indirection, and the names are defined in a preamble above; a model that reads only the granting clause of such a deed will extract two phrases rather than two names, which is at least an honest failure.
Do not anchor on position on the page. The block at the top left is often “Prepared by” and the block beside it “When recorded return to” — a law firm and a title company respectively, and neither is a party. Extractions that key on “the first name on the document” return the drafting attorney with high confidence, which is the worst combination of properties an extracted field can have. Instructing the model to quote the span it used, as in extraction prompts, makes that failure visible in review.
Names carry capacity
The party field is rarely just a name. What follows it is frequently part of the identification and cannot be dropped:
- Vesting language for the grantee. “as joint tenants with right of survivorship”, “as tenants in common”, “as tenants by the entirety”, “a married person as sole and separate property”. This determines what happens on death and is a distinct field, not a suffix on the name string.
- Representative capacity. “PAT MORGAN, as Trustee of the Morgan Family Revocable Trust dated 12 June 2014” is the trust taking title, not the individual. Store the trust name and the trust date as their own fields; the trust date distinguishes two trusts with the same family name. The same applies to “as Personal Representative of the Estate of” and “as Attorney-in-Fact for”.
- Entity form. “a Delaware limited liability company” is jurisdictional information that matters when two entities share a name across states.
- Latin abbreviations in older deeds.
et ux.means “and wife”,et vir“and husband”,et al.“and others”. These are not names and they are not noise: they signal a party who exists and is not printed. Extracting “ALEX RIVERA et ux.” as a complete grantor list is wrong, and the correct output records that an unnamed additional grantor is present.
When the stamp lands on the name
The recording stamp is applied by the clerk after execution, so it physically sits on top of whatever was already printed there. On a deed whose margins are narrow, that is frequently the first line or two of the preamble — which is where the grantor’s name is. Scanned from microfilm, the result is a name with a rectangle of purple ink through it, or a name with two characters replaced by a county seal.
The failure mode to design against is not the model refusing. It is the model completing. A language model asked for a grantor name from a page reading “ALE_ RI_ERA” will very often return “ALEX RIVERA” — a plausible completion of a common name, indistinguishable in the output from a clean read. Names are precisely the field where inference is illegitimate: unlike a date or a total, a name has no arithmetic and no context that constrains it, and a plausible wrong name is a real person.
So the response is structural. Ask for the characters actually legible and a separate occlusion flag, rather than for a name:
{
"role": "grantor",
"name_as_written": "ALE? RI?ERA",
"legible": false,
"occlusion": "recording stamp overlaps characters 4 and 7",
"page": 1,
"bbox": [96, 132, 388, 156]
}
A field that a model cannot read should arrive as an unread field with a reason, never as a guess. The point of the bounding box is that a reviewer opens the page at the right place instead of hunting for it.
The acknowledgment is a second reading
There is one genuinely useful redundancy on a deed, and it is underused. Every recorded deed carries a notarial acknowledgment near the signatures, and the acknowledgment restates the grantor’s name in full: “On this 3rd day of April, 2019, before me personally appeared ALEX RIVERA and JORDAN RIVERA, known to me to be the persons described in and who executed the foregoing instrument”. It is a second, physically separate rendering of the same name, at the bottom of the page rather than the top, and therefore almost never occluded by the same stamp.
Extract it as its own field and compare. Agreement raises confidence on both. Disagreement is informative in itself — it is frequently a middle initial or a suffix present in one place and not the other, occasionally a genuinely different party where a deed was signed by an attorney-in-fact, and sometimes a scrivener’s error that is exactly what a title examiner is looking for. Note that the acknowledgment covers only the signing parties, so it corroborates the grantor and says nothing about the grantee, who does not sign an ordinary deed at all. The grantee therefore has no second source and stays a single-read field.
Party names on identity-adjacent documents are one of the places where providers’ safety behaviour differs: the same page that one vision model transcribes will occasionally draw a refusal from another, and a refusal in the middle of a county-scale backlog is an operational problem rather than a content one. If you route through a gateway like Multigrid, the fallback is a routing rule rather than a second integration — one key, one request shape, and the retry on a different provider shows up in the same per-request cost log as the original.
The recording stamp is also the source of the citation for the instrument itself, which is a separate extraction with separate failure modes — see extracting recording information, and the legal description page for the other half of what makes a deed a deed.
Top comments (0)