Fidelity benchmarks
Every conversion claim on this site traces back to an automated validation run: real documents uploaded through the same API the site uses, converted, rendered, and compared to their sources. These are the numbers. They are re-measured for every release — and when a number is bad, we say so.
219
documents in the corpus
arXiv papers across 12 fields, IRS forms, RFCs, + curated feature probes
97.7%
conversion success rate
214 of 219 completed end-to-end
~94%
text retention (editable Word)
word recall vs. the source; mean loss 0.06
100%
page count preserved
214 of 214 completed conversions
0.88
reading-order score
longest-common-subsequence vs. source order
The honest part: pixels vs. editability
Editable conversion re-typesets your document, so its pixel similarity to the source is inherently moderate (mean SSIM 0.39 on the corpus) even when content is fully preserved — that is the fundamental trade-off every converter faces and most don't mention. It's why there are four Word modes instead of one: TrueLayout reconstructs designed pages (certificates, posters) to a near-identical look with editable text, and Exact look is pixel-perfect with the text kept as an image. You choose the trade-off; we label it honestly.
Methodology
- Harness drives the public HTTP API — upload → poll → convert → download — exactly as the web app does. No shortcuts.
- Output DOCX is rendered (LibreOffice) and compared page-by-page to the source: SSIM on the content region, ink ratio, page count, table and image counts, word-recall text retention, reading order (LCS).
- Corpus: 204 real-world PDFs (arXiv across 12 research fields, IRS forms, RFCs) plus 15 curated feature-probe documents.
- Defects found by this harness get fixed and re-measured — the last run root-caused a table-rendering bug that is now covered by regression tests.
Numbers from the most recent full-corpus run; re-published with each release. Try the pipeline yourself on the PDF to Word converter — including the one-click sample certificate.