Radar reference

Max PrilutskiyCEO & Co-Founder4 min read

Radar records what happens to your translations as events and checks them with measurements. This page explains every term you meet on its pages: each event type and the integrations that send it, who produced a translation, and what a measurement's answer means. In the dashboard, Radar → Reference lists the same, with your organization's own measurements, in three sections, Events, Who produced it, and Measurements, each closed until you open it.

Event types#

An event records one thing that happened to a string. Its type is subject.action: the subject says whether it concerns the source text, a translation, a suggestion, or a Lingo.dev localization request, and the action says what happened to it.

TypeMeaningSent by
localization.startedA Lingo.dev engine started a localization request.Lingo.dev
localization.completedA Lingo.dev engine finished a localization request, with its source and target text and the engine. Glossary adherence, rule adherence, and MQM run on it.Lingo.dev
source.createdA string appeared in the source language: a new key in a source file, a new Help Center article, section, or category, or a new Crowdin string.GitHub, Zendesk, Crowdin
source.updatedThe source text of a string changed.GitHub, Zendesk, Crowdin
source.deletedA string was removed from the source language.GitHub, Crowdin
target.createdA translation appeared in a target language. Edit rate, edit distance, and edit patterns run on it.GitHub, Zendesk, Crowdin
target.updatedA translation changed. Edit rate, edit distance, and edit patterns compare it with the version Lingo.dev wrote before it.GitHub, Zendesk, Crowdin
target.deletedA translation was removed.GitHub, Crowdin
suggestion.createdA translation was proposed: a suggestion added in Crowdin, or a key Lingo.dev translated in a file from the CLI, CI, or the GitHub App.Lingo.dev, Crowdin
suggestion.updatedA suggestion was edited.Crowdin
suggestion.deletedA suggestion was removed.Crowdin
suggestion.approvedA reviewer approved a suggestion.Crowdin
suggestion.rejectedA reviewer rejected a suggestion.Crowdin

A GitHub repository sends an event for each localization key a commit adds, changes, or removes. A Zendesk Help Center sends one for each field of an article, section, or category published in each language. Crowdin sends string-level activity from its webhooks. See where events come from for how to connect each.

Several events can concern one string: a suggestion, its approval, and the updated translation are three events. Filter by event type before counting translations.

Who produced it#

The By filter and column on Radar → Logs name what produced the text or the answer.

ByNameMeaning
llmAI modelA large language model, such as a Lingo.dev engine or Crowdin AI pre-translation.
humanPersonA translator, a reviewer, or a teammate committing to a repository.
mtMachine translationA machine translation engine, such as DeepL or Google Translate.
tmTranslation memoryAn earlier approved translation, reused from a translation memory.
algoAlgorithmCode that computes the result, such as edit distance, rather than a model or a person.

Measurement answers#

Every measurement answers each event it runs on. The Value filter on Radar → Logs → Measurements narrows answers by kind:

ValueMeaning
PassedA pass-or-fail measurement found the translation met its criterion. For edit rate, a person changed the translation.
FailedA pass-or-fail measurement found the translation did not meet its criterion. For edit rate, nobody changed it.
ScoredA 0 to 100 measurement gave a score, such as MQM or edit distance.
Not applicableThe question did not apply to the event, for example a glossary check on a source with no glossary terms. Left out of averages and pass rates.
Condition not metThe event is of a type the measurement runs on but falls outside its languages, engines, origins, or sample. Answered N/A without a judge and not billed.

Built-in measurements#

MeasurementWhat it answersRuns on
Glossary adherenceShare of the relevant glossary terms the translation appliedlocalization.completed
Rule adherenceShare of the applicable rules the translation followedlocalization.completed
MQMA 0 to 100 score computed from the errors a model listslocalization.completed
Edit rateWhether a person changed the translationtarget.created, target.updated
Edit distanceHow much of the text changed, as a percentagetarget.created, target.updated
Edit patternsWhich kinds of edits a change madetarget.created, target.updated

AI Reviewers explains how each one computes its answer, and how to write your own.

Next Steps#