Machine translation has an accountability problem: you get a file back and a shrug. Was everything translated? Did the formatting survive? Are the numbers still right? Lokalizeit answers with a fidelity score — a single 0–100 number per file, built from checks a machine can actually verify. Here is how it works and, just as important, how to use it.
What the score is made of
Every translated segment is compared against its source, and five component checks are combined with fixed weights into the final number. The weights reflect how damaging each failure is to a finished document.
Completeness — weight 30%
How much of the document actually made it through translation. An untranslated segment is the worst kind of fidelity failure — a learner hits source-language text mid-course — so this check carries the largest weight.
Markup preservation — weight 25%
Bold, italics, inline styling, and structural tags should survive the trip. The check compares formatting markers between source and translation, segment by segment, and scores how faithfully they came through.
Number & link integrity — weight 25%
Figures, dates, URLs, and email addresses must come out exactly as they went in. A dosage, a deadline, or a support link that changed in translation is a correctness bug, not a style choice — any segment where numbers differ or links are dropped counts against this component.
Length sanity — weight 10%
Translations naturally expand or contract, but extreme swings usually mean something went wrong — a dropped clause or runaway padding. Segments that are unusually short or long relative to their source get flagged.
Untranslated detection — weight 10%
A segment that comes back identical to its source in a different-language run probably was not translated. These are detected and counted separately from the completeness measure.
Together these cover the practical dimensions of a faithful translation: nothing missing (completeness and untranslated detection), formatting integrity (markup preservation), factual integrity (numbers and links), and a sanity check that the content preserved its shape (length). Terminology consistency — your product names and defined terms — is handled upstream by the glossary, which renders your terms deterministically rather than hoping the model remembers them.
What it deliberately does not measure
The score is a structural measure, not a judgment of writing quality. It will not tell you whether the tone suits your brand or whether a sentence reads elegantly — a document can score well and still benefit from a human read for nuance. That is a deliberate boundary: a precise, verifiable number for the things a machine can check beats a vague promise about the things it cannot.
Reading the number
In practice the score falls into three bands. 90–100: the structure survived intact — spot-check names and terminology and ship. 75–89: most of the document is fine, but some segments tripped a check; the warning list says which kind, so review those areas first. Below 75: a meaningful share of segments has structural issues — open the review view, work through the flagged segments, and re-export.
Using it to target human review
The score's real job is triage. Reviewing every segment of every file in every language does not scale; reviewing the flagged segments does. Alongside the number you get a warning list naming each issue type and how many segments it affects — so a reviewer opens the file already knowing where to look. And because the score recomputes whenever you edit a segment, it doubles as a progress meter for the review itself.
The full methodology, including the advisory warnings for PPTX overflow and right-to-left targets, is documented on the fidelity page. To see where the score fits in a real workflow, read how to translate a SCORM file, or start from SCORM translation, document translation, and pricing.
Frequently asked questions
- Is a fidelity score of 100 a perfect translation?
- No — it means the structure survived perfectly: nothing missing, formatting intact, every number and link preserved. Word choice and tone are outside the score by design, so a spot check for terminology is still worth it before shipping.
- Why doesn’t the score measure meaning or style?
- Because a machine grading its own meaning is circular, and a vague quality grade invites false confidence. The score sticks to properties that are objectively checkable, and leaves judgment to the human review it helps you target.
- Does the score change when I edit segments?
- Yes. It is computed when a translation finishes and recomputed whenever you edit a segment, so it always reflects the current state of the document — you can watch it move as you work through the warning list.
- Does every file get a score?
- Yes — every translated file gets its own 0–100 score and warning list, whether it is a standalone document or part of a SCORM course package.