Write acceptance criteria before testing a PDF extraction

Start from the intended task

Decide the required fields, checks and finish line first so the trial gives you a clear reason to buy or move on.

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.

Set explicit behavior for unsupported inputs and unresolved fields. A useful acceptance plan can require a visible exception rather than a guessed value. The plan should describe the deliverable and its limitations, not promise perfection across every possible document.

Copy an acceptance matrix

This matrix is an original planning example, not a reported product benchmark. Replace its fields with those required by your workflow and keep the difficult cases. The expected result should be written before running the extraction so it can reveal an error rather than merely describe the output.

UK Government: Data Quality Framework ties data quality to intended use. Use that principle to prioritize requirements that matter to the recipient.

Turn requirements into checks

1. Choose a small representative document set with page continuations, unusual notation and known limitations. Record source versions.
2. Create expected rows independently from the extraction output. Specify exact comparison rules for identifiers and justified rules for numeric normalization.
3. Define field, row and document failure categories. Decide which failures block use and which can be shared as explicit exceptions.
4. Record the review scope and release decision for each output revision. Retest affected requirements when the source or extraction process changes.

scikit-learn: Metrics and scoring explains that evaluation metrics should match the application. A single overall percentage may hide the specific failures that your acceptance matrix is designed to catch.

Avoid criteria that are too easy to pass

“File created” checks execution, not content. “Most cells look correct” lacks a denominator and scope. “No errors returned” can miss omitted pages or wrong row associations. Replace these with observable outcomes linked to source evidence.

Do not lower a requirement after a failure without recording why the intended use changed or an exception is acceptable. Keep the original result visible so the revised decision is reviewable.

Use the plan to evaluate a workflow

Acceptance means each required check has a recorded result and every unresolved issue has a visible disposition. The matrix should let a second person reach the same release decision from the evidence.

For a ParseMill buying decision, adapt the matrix to invoice supplier, invoice number, dates, amounts and line items. Test your own invoice PDFs against that expected workbook, including review and correction effort. General sample and measurement fields are requirements for a separate report-workflow discussion.