Measure

Examine what translations mean, find recurring quality problems, and understand how much content the evidence covers.

A translation can read naturally and still change what the author meant. A precise term can become ambiguous, a possibility can become a promise, or a condition can disappear. Quality measurement makes those differences available for examination, including in languages the team cannot all read.

Lingo.dev supports defined checks on translations produced here or elsewhere. A check might examine terminology, a writing requirement, or the changes between a draft and its edited version. Its answer becomes useful when the reader can see what was examined, what the answer means, and which translations it describes.

Begin with what needs to remain true #

The source text carries more than words to replace. It tells the reader what happens, under which conditions, and how confidently to expect it. A meaning check needs to examine those relationships. Finding the same product name or number in both versions is not enough.

For example, the difference between may and will can turn a possibility into a commitment. A terminology check and a meaning check ask different questions about that sentence. Both can be valuable, but one cannot substitute for the other.

Choose the question that matters for the content. The evidence follows from it:

Quality questionWhat the check needs
Does the translation preserve the source's meaning?The paired source and target, with context for any ambiguity.
Does it use a product term correctly?The term, its intended meaning and usage boundaries, and the text where it applies.
Does it follow the intended voice?A specific voice requirement and the audience or situation it serves.
How much did a reviewer change?The actual draft and the edited text, paired correctly.

The last question measures editing activity. It does not, by itself, explain whether those edits corrected errors or improved the writing. Read the definition before treating any measurement as a quality score.

Choose a way to run the check #

The starting point determines how Lingo.dev receives the evidence and where the answer appears. Measurement does not require moving translation production into Lingo.dev.

Starting pointWorkflow and result
CrowdinConnect Crowdin and select event measurements. Evidence and results appear in Radar, subject to feature access.
Local filesRun lingo score with a scorer and model configuration. It writes a local result file, without adding it to Radar.
Lingo.dev requestsTurn on automatic review for the relevant localization engines, or select engines and language pairs in a measurement's condition. Answers appear in Radar with each completed localization.

For another external workflow, first establish a supported way to supply the text. Collect translations describes the available intake and the information to retain. Defining a custom check does not create a generic upload endpoint.

For collected events, built-in measurements provide defined checks such as glossary adherence and edit distance. Custom measurements express an organization-owned criterion. Configure the selected check before generating test activity. To examine events already collected, use historical evaluation. It can measure only events without an answer, or measure the interval again; save the definitions and results needed for a comparison first, because measuring again replaces earlier answers for that measurement in the chosen interval.

Examine one answer before relying on many #

Keep the evaluated text beside the criterion. Read any explanation or finding and compare it with the words it describes. A failure should give the team a specific question to resolve: did the wording change the intended action, did the source leave that action unclear, or did the evaluator misunderstand an acceptable translation?

Make sure the answer belongs to the version being inspected. If someone edited the target after evaluation, the earlier answer describes the earlier text. Likewise, a check that could not answer is different from a check that found no problem. Results explains how the available records distinguish answers, missing evidence, and unsuccessful evaluations.

One inspected case helps establish that a check is understandable. Before relying on it across a collection, test it against known errors and acceptable translations, including natural paraphrases. A check that repeatedly flags good writing creates work; a check that overlooks a consequential change leaves the question unanswered.

Understand the extent of the problem #

The next question is how often a problem occurs and which languages or content types it affects. Report that alongside how much work was examined. A high pass rate says little about a language for which few translations received an answer.

Continue with Design a translation-quality measurement to define the criterion, test the evaluator, and calculate an interpretable result. The method helps turn individual findings into evidence the team can use for review priorities, provider comparisons, and changes to its translation process. Those decisions remain with the team.