TABLEPROOF

How we test scanned tables

We preserve the source scan, original export and settings so readers can inspect a claim.

Three evidence levels

• Documented: a statement supported by an official product source.
• Measured: an observation with a dated input/output pair and settings.
• Unverified: a capability, cost or outcome we have not established.

Inputs and repeatability

Our first corpus contains 30 fictional image-only PDFs in six categories. Variants reuse a small number of fixture families; they are not independent real-world businesses. Pristine PDFs are references, not OCR inputs. New parsing rules must be checked on fixtures not used to design the rules.

What we measure

SKU text presence, row alignment and numeric accuracy are separate questions. We retain malformed numbers, duplicate or missing rows and conversion failures. We do not correct an export from the answer key, turn skipped tests into scores or call HTTP success table accuracy.

Cost and uncertainty

We record observed API credits and elapsed time. Account plans, cash costs and manual correction time remain unverified until measured. One configuration on synthetic files does not establish a product-wide accuracy ranking.

New: PDF.co six-file scan pilot — raw outputs and observed errors