#Cleaning up after the decision

After a rollout every visitor already sees the applied arm – the config answers it, no deploy needed. The code still contains both paths. The task "Code aufräumen" (and experiments_cleanup_checklist over MCP) lists:

• the losing arms and the grep anchors x3-experiment:<key> / x3-experiment:<key> arm=<arm>,
• events and funnel steps that only a losing arm sends,
• which hosts still call the key, and when they last did.

Remove the losing branches, keep the applied one without the variant() call, run your tests and build, and deploy. The task resolves itself once the site has not called the key for 7 days while the targeted pages kept getting at least 200 page loads on every host that called before – on low-traffic pages after 30 days without calls. Never earlier: an unknown key answers control, which would silently revert the variant on a cached page or a host that was not redeployed. For the same reason a cleaned-up key keeps answering its applied arm for 90 days.

#Undo

Until the cleanup, "Zurück auf Kontrolle" (experiments_revert) serves the control to everyone again – audited and announced to the admins. Then remove the variant code instead.