Reports
Use Radar's overview and saved Explore reports to locate patterns in events and measurement answers, and preserve the calculation behind a recurring analysis.
Reports help move from one translation to a pattern. A team can find which languages account for most recorded work, which measurement's answers changed, or where a rule keeps failing. The next action depends on what the report counts: more events, fewer passing answers, and more tokens describe different situations.
Radar has two reporting screens. Radar → Overview summarizes the current month: its average score, words generated, today's events, and the values measurements computed for today's events. Radar → Explore charts events and pass-or-fail or numeric-score answers and saves those charts as reports for the organization. Open an individual answer in Results to inspect its text and explanation. Edit-pattern classifications are outside these reports and result lists; use edit rate and edit distance to locate changes, then read the compared text.
Hover a tile's name on the overview for what it counts. The overview's score, words, events, and Measurements today tiles and Explore require the reports entitlement, which comes with the Enterprise plan; without it, they say that reports are not on your plan.
Choose a report by its question #
| Question | Where to look |
|---|---|
| How much did Lingo.dev generate this month? | Words generated on the overview, described in Translation workload. |
| How much activity did each language, engine, or provider see? | Events in Explore, described in Translation workload. |
| Where did repository commits change localized content? | GitHub events in Radar → Logs, described in Repository changes. |
| How did measurement answers change over the month? | The score panel on the overview, described in Evaluation outcomes. |
| What did one measurement answer, by language or engine? | A Measurements report in Explore, described in Evaluation outcomes. |
| How consistently were rules or terms followed? | A rule or glossary adherence Score report, with individual explanations in Results; see Evaluation outcomes. |
Token and model-cost series for the account period are under Settings → Usage rather than Radar; Usage explains them.
Build a report in Explore #
- Open Radar → Explore. Select New report, or open Examples and choose a report drawn from your data to adjust.
- To follow one numeric or pass-or-fail measurement, pick it under Preset. The report becomes that measurement's answers per week over the last 90 days: Score with Average for a 0 to 100 measurement, or Value Bool with Average for a pass-or-fail one. A preset replaces the whole report, and the picker names it only while the report still matches it.
- Under Data, select Add data and choose what to count or calculate, then how: Count or Unique for a field, Sum or Average for Score, and Average for Value Bool. Each series reads events or measurement answers according to the data chosen.
- Under Filters, select Add filter, choose a field, and use is any of, is none of, or is empty. A filter's value list offers the most common values across all recorded history, not only the selected range.
- Under Group by, choose an optional Breakdown, with Then by for a second one. Breakdown offers only fields that every series in the report has.
- Above the chart, choose a Date range and the Time grain — Day, Week, Month, Year, or Total — and under Chart how to draw it, then select Save report.
Each option in Data, Filters, and Breakdown says under its name what it holds, and each event type offered as a filter value says what it means. A report's suggested name follows what it shows and starts with the measurement it follows, such as MQM · Score per week, and a chart's title uses the names the Data picker gives.
Choose from Line chart, Bar chart, Number, Horizontal bars, Pie chart, Heatmap, Scatter plot, and Dot plot according to the selected data and grouping. On a time chart, Drag to zoom narrows the range to whole periods; Reset zoom returns to the saved range.
The date ranges are Last 7 days, Last 30 days, Last 90 days, Last 12 months, which covers 365 days, and All time, from the first day with data. Each range ends at the end of today, so today is included while it is still in progress. Days, weeks, and months are cut in the organization's timezone, set under Settings → General; weeks start on Monday. Events and answers are placed by the event's time, not the time an answer was computed.
A breakdown limits the groups shown at once. Time charts keep the top 5 plus Other. A single breakdown without time keeps the top 10 plus Other, or starts with 50 groups for Horizontal bars, Heatmap, Dot plot, and Scatter plot; Show more adds another 50. A second breakdown keeps the top 100 combinations. Records without the field appear as Not set. Sort offers Highest first and Lowest first for a single breakdown without time.
For event series, Filters and Breakdown offer Event type, Target locale, Source locale, Engine, and Provider. Measurement series also offer Measurement, Value kind, Not applicable, Decided by, and Judge model. A report with both event and measurement series can group by the fields they share.
Keep and share reports #
A saved report stores its name and definition: series, filters, grouping, chart, and date range. It does not store data. Opening it reruns the query, so the same report can show different values later. Every member of the organization sees the organization's reports and can rename, overwrite, or delete any of them; a deleted report cannot be restored.
After changing a saved report, select Save to overwrite it, Save as new to keep the original, or Discard. Download CSV exports the table under the chart exactly as displayed, including formatted percentages, — for missing values, and the Other group. Groups hidden by the breakdown limit are not in the file.
Explain the headline's calculation #
Averaging percentages and pooling observations can give different answers. Consider an illustrative day with 1 pass from 1 answer and another with 0 passes from 9 answers. Their daily percentages average to 50%, but the combined population has 1 pass from 10 answers: 10%.
Radar uses both calculations in different places:
| Display | Calculation |
|---|---|
| Overview score panel | Each day pools that day's answers. The month headline averages the days that have a score, without weighting them by answer count. |
| Overview score change | The month's daily average up to today minus the same days of the previous month, in points. |
| Words generated | Words summed over the month, compared with the same dates of the previous month, named on the tile, such as vs Oct 1 – 2. |
| Explore Number chart | One pooled value over the whole range. Its change compares the same-length period immediately before: points for percentages, relative change for counts, and — when the earlier value is absent or zero. |
| Explore Score → Average | The mean of numeric 0–100 answers that are not N/A. |
| Explore Value Bool → Average | Passed answers divided by pass-or-fail answers that are not N/A. Answers on a 0–100 scale are outside its denominator. |
For a combined-population ratio, read a pooled value — an Explore Number chart or the query response's totals below — instead of averaging chart points. Null days are excluded from the overview's average, and a month with no scored answer shows — rather than zero.
Empty populations need the same care. Null, a displayed zero, and no observed rows can represent different states. Answer totals cannot identify eligible translations that were never selected or completed. Use Results to report completion and determinate coverage alongside pass rates.
Investigate the pattern behind the number #
For a score change, inspect examples under the same measurement, then compare content mix, languages, sampling, and guidance. An overall average can move because a difficult language received more work, or because another measurement started answering. An observational ranking can suggest a controlled model comparison; it does not establish that the model caused the difference.
For edit activity, read edit rate and distance together and verify the draft-target pairing. Fewer edits may reflect better drafts, less review, or a different acceptance policy. Edit-pattern answers are structured and are not included in score or pass-rate measures. Time saved needs separate evidence.
For operational problems, follow Runs and Requests. Keep absent, skipped, failed, and completed stages distinct. Translation completion, destination writeback, and published output are different milestones. Individual records support investigation, but do not supply a complete aggregate state-transition history.
Structural integrity findings also need their own interpretation. Compare missing keys, extra keys, placeholders, and repaired output against the expected contract. Absent integrity fields do not establish a successful check, and a structural failure is different from a changed claim.
Interpret event and ingestion timestamps using the source's actual time semantics. Crowdin approvals retain the suggestion's creation time, so the difference can include time before approval rather than collection delay alone. Events explains the date-filter consequence. Judge duration describes another interval. Keep provider-string events separate from request or job-locale events when calculating counts or delay summaries.
Compare spend and time on the same work #
Keep the accounting unit explicit: request, step, measurement answer, or account period. Several measurement answers can refer to one generation request, so attaching its full cost to every answer duplicates spend. Token totals are workload rather than money, and default-zero event fields do not prove zero generation cost.
Request duration is narrower than elapsed job time with queueing or review waits. Recorded cost per determinate answer can describe measurement yield, but it cannot establish human effort saved. Costs gives the required accounting boundaries.
Preserve a recurring analysis #
A saved Explore report keeps the question's definition, but not the reasoning around it or the values it showed. Keep that with the team's analysis: the decision it informs, source population, unit, measures, breakdowns, filters, interval, timezone, time grain, ordering, and treatment of groups outside the top-N display. Preserve measurement IDs as well as names, since names can change.
For every ratio, name numerator and denominator. Locale share uses all surviving filtered groups before hiding smaller ones. A quality pass rate uses determinate answers under one criterion and reports coverage separately. Period growth divides the difference by the previous period and is undefined when that previous value is zero.
Retain the observed result or retrieval date with the specification. A saved report or live query can return different data later, for example after a historical evaluation replaces answers in its interval. A downloaded CSV records what the report showed at that moment.
Finally, test the measurement behind the report. Validate quality findings examines whether flags are justified; measurement design examines false alarms, missed errors, and coverage. Agreement between two model judges alone cannot establish correctness.
Query pooled report values #
Send JSON with Content-Type: application/json to https://api.lingo.dev, authenticating with a server-side X-API-Key for the intended organization or an accepted bearer credential. These operations require organization membership and the reports entitlement; they do not require Owner solely to read analytics. Replace :orgId with the organization ID and keep credentials out of client code and saved report specifications.
| POST path | Measures to request and read from totals |
|---|---|
/v2/organizations/:orgId/analytics/measurement-scores | averageScore, reviews, scored |
For a whole-period score, use this body with the measurement-scores operation, replacing the illustrative interval:
{
"measures": ["averageScore", "reviews", "scored"],
"time": {
"axis": "createdAt",
"from": "2026-10-01T00:00:00.000Z",
"to": "2026-11-01T00:00:00.000Z"
}
}Read totals.averageScore with totals.reviews and totals.scored. averageScore is on a 0–1 scale; reviews counts every answer, including N/A, and scored counts those with a determinate answer. Omitted dimensions request the combined population; grouped responses also calculate pooled totals independently of their rows. The createdAt axis is the event's time.
Supported filters narrow the population: for example, filters: {"reviewerId":"…"} selects one measurement by its ID, and engineId, sourceLocale, and targetLocale select engines and languages. Use real IDs from the measurement reads. These are declared-query operations, not arbitrary SQL. The server supplies the organization's timezone and rejects a caller-supplied timezone; retain that timezone alongside explicit interval bounds. A query run later can return different data, so save the observed totals and retrieval date with the calculation.
Explore's own series use /analytics/events and /analytics/measurement-answers with the same request shape on an eventAt axis; they return rows without totals. Saved reports are available at GET and POST /v2/organizations/:orgId/reports and GET, PATCH, and DELETE /v2/reports/:id.