Try to fool it

Leetspeak, look-alike letters, spacing, reversal, a word glued into a username. Every trick that works is named in the answer; every innocent word that gets through is the point.

What the controls do

Context: What the text is. A username is split into words and read backwards; a review is read as prose.
Policy: How strict: kids, teen, standard or strict.
Mode: Whether the model may be asked. Fast never asks, balanced asks when the rules are unsure, thorough asks about anything not certain-clean.
Mask: How the censored text is written. The docs have the rest.

The playground runs on the site's own daily budget, not on a key, and its checks are never stored. For a project's own defaults and lists, the same tool is in the dashboard.