Historical evaluation
Apply the current measurement to recorded events, choose whether to keep or replace earlier answers, and account for unavailable evidence and asynchronous completion.
Historical evaluation runs a measurement over events already collected in your organization. Use it to apply a new criterion to earlier translations, to measure events imported from a provider's history, or to examine a bounded period under a revised definition. On the measurement's detail page, this is the Backfill tab.
For older provider content, first import GitHub or Zendesk history and confirm that the events you want are visible in Radar → Logs. Import progress counts provider items queued for collection; mapping and storage can finish later. Historical evaluation reads the stored event payloads. Imported history needs a measurement backfill because those events are collected without measuring them as they arrive.
Read what is already measured #
The Measured section shows how many events of the types the measurement runs on have an answer, N/A included, across all recorded time — for example, an illustrative “1,204 of 8,154 events measured”. It counts events with a stored payload, includes the measurement's event types whether it is on or off, and stays the same when you change the Backfill dates. Beneath the total, each origin — a GitHub repository, Zendesk Help Center, Crowdin project, or Lingo.dev — shows how many of its events are measured, such as “acme/web 1,180 of 7,600”, with Unknown origin for events whose record names none. GET /v2/organizations/:orgId/measurements/:measurementId/period-counts with ISO from (inclusive) and to (exclusive) returns the same events, measured, and origins for any interval.
Select the definition and period #
Turn the measurement on before starting: Start backfill stays unavailable while it is off. Its enabled event types determine which old events are considered, and each event is checked against its condition's current languages, engines, and sample. An event outside them receives an N/A answer, as in live evaluation.
Below From and To, the tab counts the period's events by where they came from: each GitHub repository, Zendesk Help Center, Crowdin project, and Lingo.dev, with Unknown origin for an event whose record names none. It counts the events a backfill reads, those of the measurement's enabled event types whose payload is stored, before conditions narrow them; an event outside a condition's languages, engines, or origins is counted there and answered N/A. The count is not shown while the measurement is off.
Choose a small period whose events you can inspect. Check that their payloads contain the source, target, draft, or context required by the criterion. For glossary and rule adherence, also preserve the localization engine's relevant current guidance: those checks read attached terms and rules when evaluation runs, rather than recreating generation-time context.
The interval selects recorded eventAt values. A Crowdin approval uses the suggestion's creation timestamp, which can precede the approval day. Inspect the event's timestamp when choosing a period; selecting only the day an approval arrived can omit that event.
Keep or replace earlier answers #
Only events not measured yet is on by default. It keeps the answers already in the period and measures only events without one; an N/A answer counts as measured and is kept.
Turn it off to measure the whole period again under the current definition. Starting then asks Measure the period again?, because the measurement's earlier answers in the period are deleted first and written again as the events are measured.
Save earlier answers before replacing them
With Only events not measured yet off, the deletion covers the measurement's answers throughout the selected interval, including event types or locales that are no longer selected. Save the earlier definition and results needed for comparison first. A skipped payload or failed replacement can leave a gap where an answer previously existed.
Start a bounded evaluation #
- Open the measurement's Backfill tab.
- Set From and To. The date controls include both chosen UTC days and cannot go past today; choosing the same day evaluates that day.
- Decide whether Only events not measured yet should stay on.
- Select Start backfill, and confirm if you are replacing answers.
- Follow the progress line — Backfilling, then Backfill finished or Backfill failed — with its count of events measured. The backfill also waits in the background jobs dock in the corner of every page; select either one to open its progress.
The underlying start contract is POST /v2/organizations/:orgId/measurements/:measurementId/backfill with ISO timestamps, inclusive from and exclusive to, requiring from < to, and unmeasuredOnly. Through the API, unmeasuredOnly defaults to false, which replaces answers; send true to keep them. For one UTC day, the interval is midnight on that day through midnight on the next; the UI converts its inclusive final date to that next-day boundary. The response is HTTP 202 with an asynchronous run id. A second backfill for the same measurement is refused while one is running, and GET /v2/organizations/:orgId/backfills lists recent runs; status=running&startedBy=me narrows it to the running ones you started.
Reconcile the new answers #
Processing reads events in pages and queues measurements only where the stored payload is available. A finished backfill means its events were measured, but each answer still records its own outcome. Backfilled runs are metered like live ones, except answers skipped by a condition. Backfilled answers do not generate localization engine suggestions.
Inspect the selected events and results. Result history does not expose the backfill-run identity, so match event identity, measurement identity, event interval, and computation time. Keep expected events with missing answers in the reconciliation rather than deriving completion from the rows that happened to return.
Read the reason for each new N/A. It may reflect a condition's selection, missing evidence, no applicable guidance, or a judge's inability to give a determinate answer. An execution failure needs different recovery from an eligible translation that lacks its source.
Preserve a comparison outside replacement history #
If the question is whether a revised criterion detects errors more reliably, keep the old and new definitions with separately adjudicated examples and their outcomes. If the question is whether translations improved, preserve both translation revisions under comparable conditions.
Backfill has no established cancellation or answer-rollback operation. Avoid overlapping reruns as a way to investigate a missing result: first identify the selected events, missing payloads, and returned answers. Compare translation models explains how to retain repeatable trials without relying on mutable history.