What is measured, how it is scored, and what the numbers do not say.
This page is generated from the text Exitproof prints in every report, so what you read here is what a customer reads beside their own results.
Every document was analysed statically, then converted by each suite twice: once to PDF and once back to .docx. The PDF is compared against the original for text, page count and - where a Word reference PDF exists - page images. The round-trip .docx is re-analysed and its feature counts are diffed against the original, which is what catches a content control or a field silently turning into plain text.
Fonts are checked twice over: which faces the document asks for and whether a metric-compatible substitute exists, and separately which faces actually ended up embedded in the produced PDF. The second is the one that reflects the suite; the first is a property of the document and is identical for every suite.
Two scores are reported per suite. The full score counts every issue, including the round-trip ones. The PDF-only score counts just what the rendered PDF and the document's fonts decide - text coverage, page count, blank pages, visual difference, error text and font substitution - and is the only one two suites may be compared on, because a suite that hands the document back unchanged can raise no round-trip finding at all.
Each suite starts at 100 per document and loses points per issue by severity. The status is RED on any CRITICAL issue or below the red threshold, AMBER on any MAJOR issue or below the green threshold, GREEN otherwise. A document's overall status is the worst of its suites.
A best suite is named only on a strict win. Equal scores are reported as a tie rather than broken arbitrarily.
The page count in app.xml is only as reliable as the application that last saved the file. A document last saved by something other than Word, or never repaginated, can carry a stale count. Where a Word reference PDF was supplied, the reference page count comes from the PDF instead and the source is named per document.
Without a reference PDF there is nothing to compare page images against, so the visual check falls back to comparing the suites with each other. That finds the odd one out; it cannot say which of them matches Word.
SSIM is computed on greyscale page renders at the configured DPI. It is sensitive to a shifted baseline or a substituted face, which is the point, but a low score on a page that is mostly one large image says little.
Text coverage is measured on word tokens. A suite's own line breaking moves tokens between lines without losing any, and a word broken by an automatic hyphen is rejoined before counting, so that it is not read as a word that went missing. The rejoin covers the break as PDF extraction reports it, a control character between the two halves. A suite that wrote a plain hyphen there instead would still have the word counted as missing; none has been observed doing so. Line breaking that coverage ignores is still visible in the page-image comparison.
Comparing suites is only meaningful at an equal font set. The stock ONLYOFFICE image ships Microsoft core fonts, which the other two images do not and may not redistribute; that image alone therefore reports no substitutions. The header of this report names the font set and its hash for each suite.
Collabora converts each document under the language that document declares, so it has no single locale to report. The other two run under a locale fixed by this tool. Hyphenation follows the locale and moves line breaks, which moves page counts.
ERROR_TEXT_VISIBLE relies on a table of the error strings Word and LibreOffice print in place of a broken field. Only the English strings have been confirmed against real output from both; the other languages are provisional, so a broken field in a non-English document may be missed, and the count for this code should be read as a lower bound.
Macros are detected, never executed, and never converted. A document whose logic lives in VBA will convert as a document and still not work as a tool.
Two suites can be compared only on the PDF-only score, and only when they ran on the same set of fonts. The full score includes findings about the converted .docx, and a suite that returns the file's own parts unchanged cannot raise one, so its full score is not comparable with the others'. Every report states the font set and its hash per suite, and marks the comparisons that are not valid.