How to use this page: Each row shows one flagged post with the member, date, category, engagement metrics, and a direct link to the original post on X. Click any row to expand and see the classifier's reasoning, the counting rule that applied, and additional metadata. The methodology used to classify these posts is documented at the methodology page. Category 4 (named threats) had only two flags total across the entire 122,106-post corpus and zero in the top-10 ranking, so it is not represented here.
EC-07 reclassification — April 27, 2026: A user-surfaced classifier inconsistency around the criminal-identity-as-class-attribute pattern was identified during state-delegation review after the initial run. EC-07 (a Cat 2 clarification covering criminal-identity nouns applied as identity attributes to ethnic, religious, immigration-status, political-affiliation, or institutional classes) was added to the methodology, and 1,239 candidate posts were reclassified by Haiku against the updated prompt. 221 posts flipped from OUT to IN (197 R / 24 D / 0 I); 12 of those flips appear in this audit trail (marked EC-07 in the row detail). Members marked ★ are small-sample entries — their 95% CI upper bound exceeds 3× the point estimate. Full methodology refinement narrative is in Section 12, Task 4 of the methodology page.