Events
Find collected translation activity, inspect its text and provenance, and follow the measurements that examined it.
An event records something that happened to content: a suggestion was added, a translation changed, or a localization completed. Open it when you need to establish what Radar received before interpreting a measurement answer.
The event type matters. A suggested translation, an approved suggestion, and an updated target can concern the same string while describing different activity. Counting all three as three translated strings would overstate the work.
Find and open an event #
Radar → Logs has two tabs: Organic lists every event received, and Measurements lists every value a measurement computed on an event. Narrow either to the Last 7 days, Last 30 days, Last 90 days, or All time, and by Languages, event Type, and Source; on Organic, By selects the producer type. Languages, Source, and By offer only the values your organization's events carry, plus any value already chosen. By names each code in full: llm is an AI model, human a person, mt machine translation, tm translation memory, and algo an algorithm. Languages matches the event's locale exactly. Times, the day each row falls under, and the window follow the organization's timezone, set under Settings → General, as the audit log does; the window counts those days through the end of today. Hover a time for the exact time with its zone. Clear drops the languages and every filter at once and keeps the window.
Hover an event type, in the Type filter, the table, or the event's page, for what it means. Reference lists every event type with the integrations that send it.
Open the event and inspect the available source, target, draft, and context. Its page shows Type, Locale, By, Actor, At, Ingested, Unit, and Correlation, then the measurement answers under Values and the stored Payload. Keep its identity and recorded event time when sharing an investigation. The read contract uses eventId and the event's eventAt value as at to locate the detail.
Event history is cursor-paginated: the default page size is 50, with a maximum of 200. Continue with meta.nextCursor under the same filters until no next cursor remains when you need the complete selected history. Visible rows on one page are only part of that history.
Read the record's identity #
| Field | What it identifies |
|---|---|
eventId, eventType | The recorded occurrence and kind of activity. |
unitId, displayKey | The underlying content unit and its readable identifier. |
locale | The locale associated with this activity. |
providerType, providerId, actorId | Available producer or actor information, whose role depends on the event. |
correlationId | An available correlation reference, or null. |
eventAt, ingestedAt | The source-derived event time and collection time. Interpret their difference using the source's timestamp rules. |
Detail payload | Stored event evidence, or null when unavailable. |
Understand the source's unit #
Lingo.dev events can describe a recorded translation request or one target locale of a document job. Request-derived text contains the source and target key-to-text objects. Job-derived text groups units by source file and includes completed target values; it can lack a generation model and token usage. Neither form is automatically one string.
Crowdin events identify a provider string and locale. Source additions, updates, and deletions; target changes; and suggestion creation, editing, deletion, approval, and disapproval are mapped to event types. Empty old and new target text produces no target-change event. Plural-form objects are retained as JSON text.
For a target change, the edit measurements compare with the string's previous version: the latest earlier target creation or update for the same string and locale, or the text before the change that the event carries. That version is usually the delivered translation, but it can be an earlier edit. Verify the pairing before attributing every difference to human correction. A repeated approval of the same suggestion is recorded once. Repeated disapproval actions use receipt time in their event identity, so retries do not have a universal deduplication guarantee.
Crowdin translation and suggestion events use the translation's createdAt as their event time, including approvals. An approval received today can therefore appear under the older suggestion's creation date. Receipt time distinguishes repeated disapprovals of one suggestion but does not replace that event time; an approval is recorded once per suggestion. Date filters and historical evaluation use eventAt; search the recorded date rather than only the approval day. The gap to ingestedAt can include time before approval, so it is not a pure collection-delay measurement. A missing or invalid source timestamp falls back to receipt time.
Provider, generator, and reviewer identities also differ. A model on a Lingo.dev request identifies generation; a Crowdin user on an approval identifies activity in that provider. Observing approval does not establish that the translation was exported or loaded by the application.
Follow the answer to its evidence #
Open the event's measurement result and read the criterion with the text it examined. For a suspected terminology failure, inspect the term and its usage boundary. For a rule failure, inspect the particular rule and explanation. Results describes the available answer, suggestion, usage, and judge trace.
An algorithmic measurement can have no AI request because it computes the answer directly. A model-based result may expose the actual prompt and response; read those when the finding appears to assume missing context. An unavailable trace supplies no additional judgment.
If an agent assists with investigation, authorize reads explicitly. Connected tools can also change configuration or submit new translations, so a reproduction requires a separate decision and may incur usage. Requests helps distinguish inspection of the original work from a new attempt.