TableProof PDF.co scan pilot
Run date: 2026-09-15 UTC (2026-09-16 Pacific/Auckland).
Six synthetic files, one first fixture from each of six categories, seven pages.
All cases reuse fixture-01 underlying line items; these are not six independent business datasets.
Input scans: https://gettableproof.com/samples/ (image-only PDFs, not pristine originals).
Endpoint: /v1/pdf/convert/to/csv, lang eng, OCRMode Auto, OCRResolution 300, inline true, async false.
One attempt per file. No template, crop, manual correction or answer-informed parsing.
Original CSV response body is retained, including blank columns, metadata and errors.
run-record.json is a public metadata extract. Account data and temporary signed links are omitted.
A request returning HTTP 200 is not evidence of a usable table.
Exact SKU text presence searches all parsed CSV cells with ASCII alphanumeric/hyphen boundaries. It does not measure field accuracy, row alignment or numeric correctness.
No normalization, Unicode substitution, or correction is performed. Record numbers are one-based parsed CSV records.
The separate table scorer was not applied: these raw exports are not one stable eight-column schema. No zero score is substituted for that limitation.
Total observed credit balance reduction: 238. Conversion: 196, upload: 42.
No new subscription or credit purchase was made. Actual account plan, service version, and dollar cost per credit are unverified.
Local elapsed time includes upload, conversion and balance checks; no throughput claim follows from one run.
Source code is supplied for reproducibility; never put your API key in a public site.
Raw and run files are original test evidence; the selected error examples in the article are not an exhaustive count.
