Reading APNs from document images: accuracy without wrong notices

A notice under Section 27297.7 starts with a parcel. Without the assessor's parcel number (APN) there is no owner of record, no tax-bill address and no notice. Many Recorder indexes carry the APN in a field, but not all of them do, and even where they do, the field is sometimes blank or mistyped. The APN is still usually there on the recorded document, often on the first page, because many counties require it.

SentraDeed can read it from the document image. The design goal is not "read as many APNs as possible." It is never send a notice to the wrong parcel.

Where it looks

Reading starts with page one, since that is where the statute lets a recorder require the APN. If page one has no usable APN, SentraDeed moves on to later pages, up to the first ten, because some documents carry the parcel number on an exhibit or legal-description page instead.

How a reading is scored

Every candidate gets a confidence score built from three things:

• How clearly the characters were read. The OCR engine's own confidence for the characters in the field. Where the engine can re-read the field, it does so several ways and checks that the reads agree.
• Where it was found. Text right next to a label such as "APN" or "Assessor's Parcel No." scores higher than an unlabeled number that merely looks like a parcel number.
• Whether it fits the county's format. A value that matches the county's APN pattern as printed scores higher than one that only fits after swapping a letter O for a zero.

Each attempt is stored, along with a cropped image of exactly what was read.

When a reading is accepted automatically

A reading is only accepted without a person when all of these are true: the confidence is above the county's threshold, every APN read is on the Assessor roll, and the person who signed the document matches an owner of record for that parcel. If any one of them fails, the reading is not used on its own.

When it goes to a person

Everything else, whether low confidence, a number that isn't on the roll, a signer who doesn't match, or two different APN-like values on the page, becomes an "APN from image needs review" exception. Staff see:

• The page of the recorded document with each APN reading outlined where it was found.
• A crop of each reading, the page it came from, the value read and its confidence.
• The owner of record for each candidate parcel and the signer on the document.

Staff confirm the right reading, correct it, or enter the APN, and the document is matched again from there. The decision is recorded in the audit log.

Why an OCR'd APN never goes to signer review

One design choice deserves a note. Normally, when a signer doesn't match the owner of record, SentraDeed still mails the notice and flags the case for fraud review. When the APN came from an image, that rule would be dangerous: a misread digit and a fraudulent signer look exactly the same, and mailing on a misread APN would send someone else's notice to a stranger. So a mismatch on an image-read APN always goes to a person before any notice is created.

Measuring it

SentraDeed is developed against a synthetic set of document images with a known answer for every one, including blurred and rotated scans, missing labels, APNs on later pages and documents with two parcels. The image-mode scorer runs OCR on every document and compares the result with the answer key, so any change to reading or scoring is measured before it ships. On that synthetic set, image-mode OCR has produced zero notices to the wrong parcel. Live county images will differ, which is one reason onboarding includes a parallel run.

An optional AI second reader

SentraDeed also includes an optional AI second reader for APNs. It is optional and off by default; a county turns it on. On the same synthetic test set (98 deed images across 7 scan variants) it read 98 of 98 APNs correctly, invented none on pages without an APN, and led to 0 wrong-parcel notices through the matching rules. It can only confirm an OCR reading it matches exactly; it never decides alone and never clears fraud signals. For a county, the recommended rollout is shadow mode, then suggestions, then corroboration, after a trial on the county's own images.

No OCR is perfect, and the system does not pretend otherwise. The point of the design is that imperfect reading costs staff a few seconds of review, not a notice in the wrong mailbox.