Annotate a batch of experiment runs

Write human annotations to a batch of runs in an experiment.

Idempotency: Writes use upsert semantics — submitting the same annotation config name for the same run overwrites the previous value. Retrying on network failure will not create duplicates.

Unmatched record IDs: If a record_id does not correspond to an existing run in the experiment, the annotation for that record is silently ignored. The response will still include an entry for it. No error is returned.

Payload Requirements

• experiment_id is the path parameter for the target experiment.
• annotations is a list of per-run annotation inputs, each identified by record_id.
• Annotation names must match existing annotation configs in the experiment’s space.
• Up to 500 runs may be annotated per request.

Valid example

{
  "annotations": [
    {"record_id": "run_abc", "values": [{"name": "quality", "label": "good"}]}
  ]
}

Invalid example (annotation name not found in space)

{
  "annotations": [
    {"record_id": "run_abc", "values": [{"name": "nonexistent_config"}]}
  ]
}