Results

Read a measurement answer with its text, criterion, reasoning, and execution evidence, then account for unanswered work.

A result connects a measurement with the text it examined. To use it, open the answer beside the criterion and the evaluated revision. A finding about a removed condition should identify the condition and the target wording; a number alone cannot establish that change.

For example, an edit-rate answer of true means the draft and target differ after whitespace normalization. The same boolean direction in a custom requirement might mean the translation passed. Read the definition before interpreting the value.

Open the right history #

In Radar → Logs, the Measurements tab lists every answer, newest first. Narrow it by Measurement, Languages, Type, and Value — Passed, Failed, or Not applicable — then open an answer for its Measurement result. An event's own page lists its answers under Values. The result-history contract supports time, event, measurement, locale, event-type, and verdict filters. Boolean pass/fail filters select boolean answers; they do not impose a threshold on numeric scores. N/A selects answers marked not applicable.

The result-list contract defaults to 200 rows and allows up to 500 per page. Follow meta.nextCursor with the same filters to read the complete selected history. Detail lookup uses the event ID, measurement ID, and event occurrence time as at.

For a translation generated in Lingo.dev, the answers belong to its localization.completed event. Match the event's job ID and target locale in Radar → Logs → Measurements. Review Lingo.dev translations explains which completed work can receive measurements. Open the job in Studio → Localization API to inspect its model requests.

Read the answer fields #

FieldInterpretation
valueKindbool or score, identifying the answer scale.
valueBoolean true as 1, false as 0, or a score from 0 through 100.
notApplicableWhen true, the row has no determinate boolean or numeric answer. Read this before treating value as a score.
reasoning, suggestionAn available explanation and proposed correction. Either can be null.
eventAt, computedAtThe source-derived event time and the time this answer was computed; event time follows the provider's timestamp rules.
Producer fieldsThe recorded evaluator or algorithm, rather than necessarily the translation's generator.

The stored numeric value for a null answer is zero with notApplicable: true. That combination means N/A, not a zero-quality judgment. A genuine determinate zero has notApplicable: false.

Detail adds token usage, costUsd, and an available aiRequest containing model, timestamps, input, and raw output. The result page shows them as Answer with Reasoning and Suggestion, the Translation examined, and Judge request with what was Sent and Replied, beside Judge, Took, Tokens, and Cost. An algorithmic answer needs no AI request. Older or unavailable request evidence can also leave this field null. A null explanation on an otherwise valid answer is not itself an execution failure.

Read the judge input when its explanation seems to depend on missing text or guidance. Evaluate a suggestion against the source before applying it; a suggested correction has not become accepted output.

The detail contract includes a structured answer field, but current result queries return only boolean and numeric rows. Stored JSON edit-pattern labels are not available through these list/detail paths. A backfill that measures a period again also replaces prior answers, so result history is not an immutable archive of criterion versions.

Interpretation #

N/A combines several situations that require different follow-up. The event may fall outside the measurement's languages, engines, or sample, with reasoning that begins “Condition not met”. A rule may genuinely not apply. A selected translation may lack its source. A judge may run and still be unable to decide. Use the reason, payload, and execution evidence to distinguish them rather than assuming every N/A is a completed judgment.

Observed situationWhat to do with it
Applicability is unclearResolve whether the criterion belongs in this assessment; keep the uncertainty visible.
The criterion does not applyExclude it from that criterion's quality denominator and retain the reason.
An eligible translation lacks required evidenceKeep it in the eligible population and record the evidence gap.
The judge completed but gave no determinate answerCount the completed evaluation separately from determinate answers.
An eligible translation received a passing or failing answerInclude it in the determinate denominator; include only passes in the pass numerator.
No completed evaluation is establishedKeep the missing or failed work visible and investigate its last known milestone.

These are analytical distinctions, not a promise of separate UI states. The selected population comes from your workflow record; returned answers cannot identify every eligible translation that never reached evaluation.

For an illustrative population of 30 eligible translations, 18 passes among 20 determinate answers give a 90% pass rate and determinate coverage of 20/30. If another five evaluations completed without a determinate answer, completion coverage is 25/30. The other five remain without a completed evaluation. Report those quantities separately so the pass rate cannot conceal missing evidence.

Sampling is a deliberate selection step. Events omitted by sampling are not execution failures, but they still affect coverage of the full eligible population. Design a translation-quality measurement explains how to define that population and test the criterion.

Missing and failed results #

Start with the original event or request. If it is absent, inspect collection or generation. If it exists, inspect the selected measurement, its enabled conditions with their languages, engines, and sample, and the required evidence.

For event processing, recoverable judge or storage failures are retried; malformed messages and defects can be dropped. A visible answer list is not a complete attempt ledger. Successful translation also does not guarantee that a measurement was selected or queued for it.

Use Events for collected evidence, Requests for model input, output, and errors, and Runs for workflow progress. A human-review wait belongs in Inbox. A failed destination write is separate from a translation or evaluation failure. Repeatedly generating translations to obtain one missing review can create unrelated work without resolving the original gap.

Act on the evaluated revision #

Confirm that the target being inspected is the evaluated version. For professional review, compare the returned text with what was actually sent for review, which may already include rephrasing. A late vendor result or retained review fee does not establish that the team accepted that text.

After validating the finding, correct the target in its owning workflow or clarify the source, glossary, rule, or evaluator responsible for the problem. Recheck the changed revision. Validate quality findings develops that investigation; Inbox reviews covers the review decision itself.