Skip to main content

Get Run Analysis Endpoint

GET 
/runs/:run_id/analysis

Return pre-aggregated scorer statistics and histograms for a completed run.

Provides per-scorer mean, percentiles, and 10-bucket score distributions. Designed to drive the UI output analysis view (histograms, metric table, failure threshold counts). For row-level data use GET /runs/{id}/results.

Optional group_by accepts an embedded categorical column (e.g. scenario) or a qualified dataset dimension ID (e.g. declared:scenario or derived:image_properties.brightness) and adds group_stats to the response for breakdown views. Qualified dimensions also return join coverage. Derived callers may pin expected_definition_version; stale definitions return a structured 409. dimension_columns lists only grouping columns embedded in the run artifact. Discover qualified IDs through GET /datasets/{dataset_id}/versions/{version_id}/dimensions.

Optional slice_metric requires group_by and names one supported full-dataset metric result column. The metric is recomputed from the rows inside each dimension group and returned in slice_results.

Optional classification_group requires group_by. The response keeps run-level threshold and dimension clusters, but replaces per-class image classification metric clusters with those from the selected group. group_counts reports every group, including groups with no class failures.

Registered class failure attribution remains part of the standard analysis response. include_classification_failure_clusters additionally opts into reconstructing failures across every embedded dimension. Both paths read classification row payloads; the flag avoids the additional all-dimension work for callers that do not render the full Failure clusters view.

A trace_metric_invoke run reports a DIFFERENT shape, because it can measure one metric at several targets at once and a metric name alone then identifies nothing. Instead of the wide path's scorer_stats/score_distribution/metric_metadata maps it returns series: one entry per (metric, partition_type, partition_id) measurement, carrying its identity, its stats, distribution, rollups, session_scopes and occurrence_count. rollups holds session/system summaries derived from the stored verdicts, each naming the reduction and version that produced it, and present only where the metric declares that reduction admissible over that granularity — session_scopes lists the sessions regardless. clusters and group_stats reference an entry by its opaque series_id. Pooling two measurements would yield a mean weighted by trace shape rather than quality, which is why they are never merged. Non-trace runs are entirely unaffected.

Returns 404 if the run does not exist, or if it has neither completed nor failed with output already written. Returns 422 if group_by is not a valid dimension column. Returns 424 if a referenced dataset dimension's own data is currently unavailable and must be recomputed.

Request​

Responses​

Successful Response