OCR read the liability cap as $100,000 instead of $1,000,000 - a comma misread and the agent never flagged the number...

TL;DR

OCR turns $1,000,000 into $100,000 and nobody blinks if you quote the number as fact. Run two OCR passes and compare, sanity-check the magnitude against related numbers in the same document, and mark any disagreeing amount as uncertain instead of reporting it. One comma is not a rounding error.

The query

OCR read the liability cap as $100,000 instead of $1,000,000 - a comma misread and the agent never flagged the number looked wrong

Use this when

• An agent reports dollar amounts extracted from scanned contracts
• Extracted caps, fees, or totals look plausible but have never been checked
• Your pipeline quotes OCR'd numbers as facts with no uncertainty flag

Not for

• Table structure extraction from scanned schedules
• Clause classification or labeling
• Amounts in clean digital PDFs with selectable text

Steps

1. Never trust one OCR pass for money

Run two engines (or the same engine at two DPI settings) over the pages that contain amounts, and compare the extracted numbers.

Expected output: agreement on every amount, or a flagged discrepancy with both readings attached.

2. Add a magnitude sanity check

Compare each extracted cap against related numbers in the same document: total contract value, annual fees, other caps. If an amount is off by roughly 10x from what the document's own context suggests, flag it.

Expected output: a $100,000 cap sitting next to a $1,000,000 total raises a flag instead of sailing through.

3. Re-OCR the number region at high DPI

Crop the page image to the amount's bounding box, render it larger, and OCR just that region. Or route the crop to a human spot-check queue.

Expected output: the true digit count confirmed independently of the full-page pass.

4. Attach the source snippet to every extracted amount

Store the page number, the raw OCR text of the region, and the engine used alongside each amount in the report.

Expected output: a reviewer sees "$100,000" came from a specific page region and can verify it in seconds.

5. Fail loud on disagreement

If engines disagree or the sanity check trips, report the amount as uncertain and needing review. Never report a contested number as fact.

Expected output: "uncertain - needs review" instead of a wrong number in the final report.

Variant phrasings

OCR dropped a zero on a contract liability cap

Same pipeline. The magnitude check in step 2 is what catches dropped zeros specifically, since they shift the amount by exactly 10x.

agent reported a wrong dollar amount from a scanned contract

The systemic fix is steps 4 and 5: provenance on every amount and loud failure on uncertainty. The agent should never state an unverified number as certain.

Why it happens

Commas and zeros blur together on low-quality scans, and OCR picks the reading with the best local confidence, not the one that makes sense. Agents that pass the extracted number straight through with no sanity check turn a character-recognition error into a contract term nobody questions.

Edge cases

• International number formats: 1.000.000 and 1,000,000 mean the same thing in different locales. Normalize per the document's locale before comparing magnitudes.
• Amounts spelled out in words versus digits: when "one million dollars" and "$100,000" disagree in the same clause, prefer the words and flag the mismatch for a human.
• Currency symbols OCR'd as letters (S for $): keep a symbol-confusion map so the sanity check does not trip on its own false alarm.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_RBvrQ6PPXFSmUT-H0Lvrzg