AN EXPERIMENT BY GUILIN ZHANG

The numbers add up.
Do they match the page?

An independent visual check for document automation. Explore real outputs from MiniCPM-OCR and OCR Jev, then decide what deserves a second look.

Inspect the evidence ↓

01 / FOLLOW THE EVIDENCE

One source. Two different questions.

Fixed payment-region crops · synthetic invoices

Dataset Case

1. SOURCE PIXELSFull invoice ↗

The verifier receives this image, the candidate text, and its rendered image.

2. CANDIDATE TEXT

Inspect the exact candidate rendering

Candidate text as rendered for the verifier

STRUCTURAL RULES

VISUAL CONSISTENCY SCORE

A 0–1 model score for agreement with the source pixels. Higher means stronger predicted agreement; it is not a calibrated guarantee.

FROZEN REVIEW POLICY

Open the audit trail: labels, model output and timing

02 / TEST THE TRADEOFF

More AI also means more review.

80 held-out regions · 3 critical-field errors

The frozen operating point

Policy	Reviewed	Errors caught	Correct items reviewed
Rules	3 / 80	3 / 3	0
Rules + verifier	16 / 80	3 / 3	13

In this pilot, rules caught every natural error. The verifier added 13 correct items to the review queue. Its extra value needs evidence on harder, representative data.

Try a different threshold

Exploratory view of the same held-out records. The published result uses the frozen calibration threshold.

Review below

sent to review

errors caught

errors accepted

A different result on deliberate stress tests

Rules missed all 24 internally consistent currency and amount substitutions. Adding the verifier routed all 24 to review. These are designed probes, reported separately from natural OCR errors.

12 currency substitutions + 12 amount multiplications

03 / FROM READING TO WORKFLOW

Keep visual correctness and business checks separate.

Read

MiniCPM-OCR 1.0
1.3B parameters
Image → text

→

Verify

MiniCPM-OCR Jev
2B parameters
Image + text + render → score

→

Route

Rules + threshold
Keep evidence attached
Human review when needed

→

Check context

Purchase order, receipt,
supplier and policy
Business validation

Correct OCR can still fail a business check.

Supplier-invoice intake is a useful application for this pattern. Workday describes OCR, validation and exception handling in its public overview of automated invoice processing ↗. This is a personal experiment using synthetic data.