How do I send feedback about a task answer?

Send POST /v1/feedback with the answer’s trace ID, a score, and an optional correction using a key with feedback.write. A correction must match the task’s answer shape. Review and promotion into a new dataset happen separately; submitting feedback does not change the live release.

bash

curl -X POST https://api.useorlo.com/v1/feedback \
  -H "Authorization: Bearer $ORLO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"inference_log_id": "'$TRACE_ID'", "score": -1, "correction": {"category": "fraud", "priority": "high"}}'

Field	Meaning
`inference_log_id`	The `trace_id` from the run response or the `x-orlo-trace-id` header
`score`	`1` right, `0` neutral, `-1` wrong
`correction`	Optional. The answer that should have been returned, in the task’s answer shape; a correction that does not fit is refused with the field named

What happens next

A correction is pending until a reviewer looks at it in the console. A reviewer stages or rejects it. An admin promotes the staged corrections of a task together into a new dataset version with POST /v1/tasks/:task_id/feedback/promote; send base_dataset_id for the previewed dataset and expected_feedback_revision for the accepted correction set. A changed dataset returns 409 DATASET_BASE_CHANGED; a changed correction set returns 409 FEEDBACK_CHANGED. Reload and review before retrying either conflict. The next comparison runs on the new version.

GET /v1/feedback lists feedback with its status; feedback.read is enough for that. A key with feedback.write can also stage or reject corrections through the API. Give that scope only to a trusted feedback integration. In the console, the Reviewer preset permits these decisions; promotion into a dataset remains an admin workflow. A runtime key performing promotion needs datasets.write.

Review protected values

Corrections preserve the task’s answer shape, including ordinary fields named token or session. Known organization credentials and recognizable secret values are removed before saving. If that changes a correction, the response includes redaction_warning; inspect the saved correction before accepting it. If the protected value no longer fits the output schema, Orlo refuses it so you can supply a safe example value.