Record omitted fields separately from invented fields

Check for absence and invention separately

Count them separately so your trial reveals both correction effort and misleading output.

In ParseMill, select an extracted invoice field to inspect its source, correct the value and export the reviewed workbook. Try your own invoices as well as the interactive sample. Compare upload, checking and correction time with your current way of building the same spreadsheet.

Use explicit states rather than one mismatch flag. Useful categories include exact match, omitted, incorrect, unsupported addition, source blank and unresolved source. Keep the raw source evidence with the classification so the distinction can be reviewed.

Use an expected-versus-observed matrix

This invented fixture separates omissions from additions even when both occur in the same column. A filled cell is not necessarily more complete in the useful sense: it may contain a value the source never supplied. The blank source state is itself information to preserve.

UK Government: Data Quality Framework treats completeness and accuracy as distinct dimensions. Keep those questions separate when diagnosing extraction defects.

Compare in both directions

1. Start with a source-based expected field list and stable row identifiers. Include expected blank or unstated states rather than skipping them.
2. For every expected field, locate its observed counterpart and classify missing or changed content. Check qualifiers and units as part of meaning.
3. For every observed row and field, identify its source support. Flag extras that have no expected counterpart or no documented derivation.
4. Record corrections by category and rerun the comparison. Keep unresolved source ambiguity separate from a confirmed parser error.

scikit-learn: Metrics and scoring describes metrics that distinguish different prediction outcomes. For your register, define categories and denominators explicitly instead of combining all defects into an unexplained accuracy percentage.

Avoid disguising additions as normalization

A normalized unit inferred from a heading may be legitimate if the heading’s scope is confirmed and the derivation is labeled. The same unit guessed from a familiar test name is unsupported. Record the evidence and transformation that distinguish these cases.

Similarly, a generated row identifier can be useful, but it should be labeled as an internal key rather than a supplier identifier. The test should judge whether the output claims source support, not whether every added technical field is forbidden. Record generated keys in a separate namespace with an explicit purpose.

Review the corrected result

For the fixture, restore S1’s unit from source, remove unsupported source claims for S2 and S3, and correct S4 only after verifying the source. Keep the original output and correction history so the failure categories remain measurable.

For the invoice trial, include a missing required invoice field and a printed zero. Compare the extracted result and validation issues with the PDF. Track missing information and unsupported additions separately when deciding whether the reviewed workbook saves work.