What it is
Asymmetric-threshold moderation with typed Jev questions.
What it does: Scores harm and clarity, picks a category, and routes to allow, shadow or remove using different bars for false removals versus missed harm.
Why it belongs in the showcase: Makes the moderation trade-off explicit in code: removing good content is treated as costlier than leaving bad content up.
Caveats: Demo scale; no labelled eval table.
Evidence
Public MIT repo with CI badge, policy.py and a .jev question file. Four typed questions per item: harm (Score 0–4), category (Choice over harassment/spam/misinformation/violence/sexual/self_harm/benign), is_clear (Noul) and needs_human (Noul). The README example shows harassment that scores clarity 0.62 below the 0.85 removal bar and is shadowed for human review instead of removed. Thresholds are treated as explicit policy values, not hidden defaults.
Caveats
- 0★ at intake; small demo repo
- No published labelled precision/recall numbers
- Needs a TypeSafe API key
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.