Account sensitive label reapplied three minutes after support confirmed the account was clear

[taiseisouma0723-dot]

Summary

An account-level sensitive-content label has now been removed and reapplied four times. The interval between removal and reapplication has become shorter with each cycle, and every recurrence has required another Support DM.

Most seriously, X Support explicitly confirmed that neither the account nor its posts were marked as sensitive. Approximately three minutes after that confirmation, the account-level sensitive label was applied again.

No sensitive or adult content was posted during that interval. The relevant post was suitable for all ages.

Observed behavior

1. The account-level sensitive label was applied.
2. Support reviewed the account and cleared the label.
3. The label was reapplied.
4. This cycle occurred four times.
5. The interval between removal and reapplication became progressively shorter.
6. After the third review, Support confirmed that both the account and posts were clear.
7. Approximately three minutes later, the account-level label returned.
8. There was no new sensitive post that could reasonably explain the reapplication.

Expected behavior

After a successful review confirms that an account and its posts are not sensitive, the cleared state should remain durable unless new, strong, content-specific evidence is produced.

A stale, delayed, or previously reviewed classification signal should not overwrite a successful review.

Possible engineering areas to investigate

The public code includes a post-to-user aggregation path that can calculate account-level labels after an applied post-label event:

• safetyLabelToUserLevelAggregationV2Forwarder.strato⁠
• safetyLabelToUserLevelAggregationV2Processor.strato⁠
• postToUserLabelRules.strato⁠

This incident does not prove that this pipeline caused the recurrence, but possible failure modes include:

• a stale or delayed Applied event arriving after a manual clear;
• an automated writer overriding a newer human-review decision;
• successfully appealed post labels continuing to affect account aggregation;
• a race condition between support-side removal and automated reclassification;
• no durable appeal state preventing old evidence from restoring the label.

Requested fix

1. Remove the incorrect account-level sensitive label.
2. Ensure that a successful review cannot be overwritten by stale or previously reviewed signals.
3. Expose which post label, rule, model, or event caused the account label to be reapplied.
4. Add a regression test for: human review clears account → no new sensitive content → delayed automated event must not restore the account label.
5. Make successful appeal outcomes durable.

Impact

An account cannot be used reliably when an incorrect restriction can return only three minutes after a successful review. It also suppresses unrelated all-ages posts and forces users and Support to repeat the same process indefinitely.

Related earlier report: #16105⁠