Cause attribution

Cause attribution is the step from "quality dropped" to "this specific change caused it." Detection establishes that a regression happened; attribution names the change responsible. Knowing a regression exists still leaves the whole search for where it came from. That search is what attribution does.

Any regression has a finite set of candidate causes: a code commit, a prompt edit, a model change, a tool whose behavior shifted, or an upstream system the agent depends on. Attribution is an investigation over that set. You gather the failing traces into one cohort, line up the onset of the failure against the change history, and test each plausible candidate against the evidence in the traces until the ones that don't hold are ruled out.

Localization is what makes the search tractable. Failures cluster around the code path that produces them. When the failing traces share one path and the failure started at a known time, the candidate set shrinks from everything shipped recently to the handful of changes that touched that path in that window.

A finished attribution is a grounded explanation: the named cause, the failing traces and findings that support it, and the hypotheses that were checked and ruled out. Because the cause is a specific change in a specific place, the attribution also says where the fix belongs.