Validation should flag suspicious numbers, not invent replacements
Use automated rules to find suspicious numbers, then verify them against the source. Validation can flag an error but should not silently invent a correction.
The outcome is a layered validation process that finds suspicious numbers with rules, checks them against the source, and records corrections without silently replacing an uncertain value.
Build a review sample with Invoice and Receipt OCR. Use the next control in Table OCR Quality Control: Rows, Columns and Merged Cells and Form OCR: Checkboxes, Blank Fields and Repeated Labels.
When the numbers come from axes, labels, or plotted marks, apply these checks in the chart image to reviewed data workflow.
Numeric validation controls
- Retain the raw OCR token
- Define format and range rules
- Recalculate arithmetic independently
- Record every correction
- Review consequential fields even at high confidence
Layer format, range, arithmetic, and source checks
Preserve the visible token
Store the raw OCR token before normalization. Keep leading zeros, separators, signs, currency, units, and surrounding label text so a reviewer can understand what was seen and what was changed.
Apply field-specific checks
Dates can be checked for valid ranges, totals can be recalculated, identifiers may have length or checksum rules, and percentages can be compared with expected bounds. A passing rule does not prove the value is correct.
Distinguish correction from inference
Changing O to 0 because a field is numeric is a constrained correction only when the source supports it. Record the original token, corrected value, rule, reviewer, and source location.
Review high-consequence fields regardless of confidence
Average confidence can be high while one account number or decimal is wrong. Define mandatory review fields from the business consequence rather than from engine score alone.
Use rules to flag numbers, then return to the image
A date can pass a format check and still describe the wrong day. A total can balance after two compensating OCR errors. Use layered checks: allowed characters, length, range, checksum where defined, arithmetic relationships, and consistency with nearby labels.
A failed rule should create a review task, not an automatic replacement. Compare the value with the source, record the raw OCR token, enter the reviewed value, and state why it changed. Apply mandatory source review to identifiers, dates, amounts, dosages, units, and any other number that drives an action.
- Validate format and relationship
- Never infer a replacement silently
- Review consequential numbers regardless of confidence
Validation layers
| Layer | Example | What it proves |
|---|---|---|
| Format | Date matches YYYY-MM-DD | Shape only |
| Range | Month is 1 through 12 | Plausibility only |
| Arithmetic | Subtotal plus tax equals total | Internal consistency |
| Source review | Reviewer checks the visible glyph | Agreement with the supplied image |
Rules detect suspicion, not truth
A value can pass type, range, and arithmetic checks while still being the wrong source number. Validation must not silently replace an uncertain token with the value a formula expects.
How the evidence supports layered checking
The references explain OCR measurement and data-quality concepts. Direct source comparison remains the final check for consequential amounts, dates, units, and identifiers.