Five thousand synthetic abuse reports, six months of queue history, and the audit trail behind every decision. Nothing here is real — the shape of it is.
A working model of the thing a Trust & Safety team actually operates: reports arrive, get triaged against a policy, get actioned by a moderator, and sometimes get appealed. Volume is dominated by spam and harassment. Child-safety reports are rare and nearly always critical. Response targets tighten as severity rises, and they get missed more often at the bottom of the queue than the top.
Not everything gets handled. A persistent backlog of 426 reports sits unclaimed — heavily weighted toward low severity, because that is where queues actually rot. Most of it is already past its response target.
No raw reported content is stored here, by design. Each report carries a neutral summary of the complaint, a hash pointer to the quarantined item, and the automated classifier score. Child-safety reports carry no pointer at all — they are marked for a restricted specialist surface only. This is how production analytics layers are built: queue analysis should never require exposure to the material.
All records are generated. No real user, case, or client data appears anywhere in this dataset.
The database runs entirely in your browser — no server, no login. Open a table, or write your own SQL
against reports, actions, appeals, moderators,
and policies.