Use case

Correction Capture

Capturing human corrections to agent output — text edits, action changes, reroutes — and converting them into eval cases, knowledge updates, or workflow changes.

Overview

A correction that disappears into a single ticket is feedback wasted. A correction captured at the workflow boundary becomes the next improvement.

What it solves

Closes the loop between human review and system improvement. Makes the rate at which the system needs correction a tracked metric instead of an impression.

How we build it

Every correction (edited output, changed action, reroute, reject) is captured with the original trace, the corrected version, and an optional reason. Corrections are clustered into the same forensics pipeline as bad threads; each cluster has a recommended remedy (eval case, knowledge update, prompt change, tool change).

  • Per-correction trace, edit, and reason
  • Cluster by failure pattern
  • Recommended remedy per cluster
  • Closure tracked over time

What changes when it is in place

Corrections become a stream of improvement candidates instead of noise.