Root-Cause Depth

A "root cause" is real only when you've reached a named, owned mechanism and chosen a response to it. Stopping at the first level that explains the symptom is the default failure. Do not stop on a vibe — run the stop condition below.

Stop condition (evaluate every candidate cause C)

C is terminal ONLY when it passes all three gates and lands on a branch. Throughout, "the action C implies" means the mechanism-level fix — what the owner would change so the correct state is produced — assumed applied. It is NOT the response branch you pick below when the mechanism is outside your control (Accept & record does not "fix" anything, so never evaluate gates 2–3 against it).

1. Owner named. What system/process was supposed to produce the correct state? If you cannot name it (IaC, pipeline, automation, a validation that should have failed, a person/team), you are above the mechanism — keep going. Fallback: if after a few levels no single owner is identifiable (a diffuse or emergent cause), stop descending and make "identify/assign the owner" the terminal action — i.e. Escalate.
2. Recurrence test. If the mechanism-level fix were applied, could this class of failure still recur through the same path? If yes → not terminal, keep going.
3. Red-flag test. Is the implied action "manually correct this instance"? If yes → not terminal, keep going. The root cause is why that value was wrong and why nothing caught it.

Once C names an owner and passes 2–3, classify control — this is the actual stop:

• In our control → terminal. Apply the structural fix to the owning system (e.g. add to IaC), not the instance patch.
• Outside our control → terminal for descent; you can't fix the mechanism, so branch the response (one or more, not mutually exclusive):
  • Accept & record — add to a won't-fix / known-limitations list with rationale.
  • Defend — guards on our side: validation, retries, fallbacks, and/or detection/alerting so the failure is contained or caught fast.
  • Escalate — identify and engage the party who owns the mechanism.
• Mixed ownership (partly ours, partly external) → do both: apply the structural fix to our portion AND Escalate the external portion.

The investigation stops at named mechanism + chosen branch — never at "manually fix this instance."

Anti-pattern

Anchoring on a checklist / acceptance-criterion's literal "classify the cause" wording as the stop signal. Satisfying the stated bar ≠ finding the cause that makes the fix durable. The AC is a proxy for the goal; don't optimize the proxy.

Note on severity

The depth push is uniform — always name the mechanism first. Severity affects the branch, not whether you descend: a low-blast-radius, out-of-control cause may resolve quickly to Accept & record; a shared-environment or recurring one warrants Defend + Escalate.