Measurements

Select checks for collected and generated translation events and control which events each check runs on.

A measurement is a defined check that runs when selected translation activity arrives. For example, a translation change in Crowdin or a repository can trigger an edit comparison, while a completed Lingo.dev localization can trigger a check of its terminology. The event supplies the evidence; the measurement supplies the question.

The same resource covers translations collected from other systems and translations your localization engines generate. Reviewing Lingo.dev output is a measurement that runs on completed localizations; Review Lingo.dev translations describes that use.

In Radar → Measurements, choose a built-in check or create an organization-owned custom definition. The page opens on the Built in tab, and Your measurements holds the ones your organization wrote; each lists the event types a measurement runs on, its value type, languages, how many events it has measured, and whether it is on. Its current API requires organization membership for reading, changing definitions and conditions, and starting historical evaluation; the reports entitlement controls Reports, not measurements.

Turn on a built-in check #

Open a built-in measurement and choose the events it should examine: languages, source languages, localization engines, and sample. Turn on Enabled and select Save changes. The page identifies the event types it runs on. Lingo.dev sets those allowed event types and the check's implementation; your organization controls whether it runs and which events it selects.

For glossary and rule adherence, a localization engine's Reviewers tab offers the same choice per engine under Automatic review.

Choose by the question you need answered:

QuestionMeasurement
What errors does the defined category-and-severity rubric find?MQM
Did the target change from the recorded draft?Edit rate
How much did the text change?Edit distance
Did the translation apply relevant terminology?Glossary adherence
Did it follow applicable instructions?Rule adherence
What kinds of edits occurred?Edit patterns, with the result-access limitation described in the built-in reference.

The built-in reference gives each algorithm's inputs, scale, and interpretation. A true edit-rate answer means that text changed; it is not a general quality pass.

Add a content-specific requirement #

Create a custom measurement when the question depends on your content, such as preserving a condition attached to a product claim. Give it a prompt, name, value type, trigger, and languages. The creation form enables the chosen event condition when it saves the definition.

Start with activity whose evidence you can inspect. After the first event arrives, open its answer and compare the reasoning with the source and target. Test acceptable paraphrases as well as known errors before relying on a larger collection. A valid saved definition establishes that the check is configured; calibration establishes whether it answers the intended question reliably enough for the task.

Event selection scope #

A custom measurement has a separate condition for each event type it runs on; add one with Choose an event under Runs on. A built-in lists the event types permitted for it under Runs on, which you cannot change, and has one shared set of settings for all of them. When no connected integration, and not Lingo.dev itself, sends the event types it runs on, the General tab says so under Runs on, names the integrations that would, and links to Settings → Integrations. The event vocabulary includes source and target creation, update, and deletion; suggestion creation, update, deletion, approval, and rejection; and localization start and completion.

Each condition or built-in setting narrows the events of its type:

SelectorNarrows by
LanguagesThe event's target locale. Any language selects every target locale.
Source languagesThe event's source locale. Any source language selects every one.
EnginesThe localization engine that produced the translation. Any engine includes engines added later.
SampleEvery event, or a share from 50% down to 1% of the events that pass the other selectors.

Locale selection follows the locale hierarchy rather than exact text. A list containing fr selects fr and fr-CA, and fr-CA also selects an event marked plain fr; fr-CA does not select fr-FR. Through the API, an empty or absent list is unrestricted, and * matches any locale.

Engine and source-language selection need fields the event carries. Lingo.dev localization events identify their engine and both locales; Crowdin, GitHub, and Zendesk events carry locales but no engine, so a condition limited to engines does not select them. Sampling is decided per event and measurement rather than by a new draw each time: the same event falls on the same side of a given sample when it is evaluated again.

An event of a selected type that fails a selector still receives an answer: N/A, with reasoning that begins “Condition not met” and names the selector. That answer is not billed, and it counts as measured on the Backfill tab. Keep it out of a criterion's quality denominator while still accounting for it in the population.

Selection and evidence are separate. A start event can match the configured type while having no completed target. A glossary check can select the locale while finding no applicable term. Inspect both the condition and the payload when an answer is absent or N/A.

Change future checks and preserve earlier meaning #

Edit a custom definition's name, prompt, value type, or Allow not applicable on its General tab, together with its event conditions, and select Save changes. Before changing a criterion, retain the earlier definition with results needed for comparison: a different question can produce a different trend without any translation changing.

Enabled turns all of a measurement's conditions on or off together; the change takes effect when saved. Through the API, each custom condition also has its own enabled flag. Turning a measurement off stops future selection, while work already queued may still complete. Deleting a custom definition removes that definition and its conditions; existing answers remain, listed as Unknown measurement without a definition to link to. Built-ins cannot be edited or deleted as custom definitions.

For already collected activity, use Historical evaluation, on the measurement's Backfill tab. It can measure only the events that have no answer yet, or replace the answers in the chosen interval.

Measurements run asynchronously. Their answers inform inspection; they do not edit translations or block publication. Configure translation stages and acceptance separately through Studio.