ENGINEERING NOTES / EVALUATION

English documents, put to the test.

A dated, inspectable evaluation of the current processing workflow.

29 OCT 2026 · PUBLIC BASELINE + ISOLATED REGRESSION

What we tested

Eight files across PDF, DOCX, XLSX, TXT and Markdown. Public teaching materials test ordinary layouts; original fictional files add known names, emails, international phones, an address and an organization. The PDF set contains five pages, including a rotated two-resume layout and an image-only agreement.

Baseline fileTimeObserved result
Public sample CVs (1 page)105.70 s38 detections; contact details masked, some contextual organizations missed.
Blank invoice (1 page)26.54 sNo redactions. Empty field labels preserved.
Public contract excerpt (2 pages)39.01 s2 detections. Legal prose remains readable.
Fictional TXT / Markdown6.56 / 4.92 s9 occurrences each; amount and ordinary prose preserved.
Fictional DOCX6.84 s9 occurrences; repeated name keeps the same label.
Fictional XLSX6.66 s7 occurrences; numeric cells and SUM formula preserved.
Image-only agreement30.92 s9 occurrences; visual inspection exposed clipped letter edges.

These are single baseline runs through the public API before the refinements below, including upload and polling. They are not speed guarantees or a benchmark against another product. Detection counts are not accuracy scores.

What the tests changed

All eight files subsequently completed in both modes (16 outputs). Native outputs removed the designated fixture values; repeated names, ordinary numeric cells and formula expressions were retained. Ten rendered PDF pages were checked for hidden source text. Two additional Chinese fixtures checked text and embedded-image preservation.

Excel formulas are preserved in manual calculation mode. Uncached formulas can display 0 until recalculated in your spreadsheet application; the gallery shows this actual output, not a retouched total.

What this does not establish

This is a small mixed test set, not a representative English benchmark. We have not measured population-wide precision or recall. Organization names and context-dependent identifiers can be missed or classified inconsistently. We observed false positives too: a generic “Central Government Body” term in the contract and some CV headings were masked. Small punctuation may remain in scanned phone fields. Handwriting, degraded scans and complex tables need broader evaluation.

Sources and reproducibility

Sources retrieved on 29 October 2026. Public source documents are linked, not republished here. Downloadable gallery files are original fictional fixtures. Office previews are rendered from the actual files; the downloads retain their original file formats.

Inspect original and result →Run the REST demo →