Moderate Realtime Turn

Screen one turn of a live call. Returns whether to end the call.

Body: text: the transcript of a single turn (<= 1000 chars) source: "child" or "panda" — logged as a counter only, never stored

Returns {"flagged": bool, "action": "continue" | "end_call", "categories": [...]}.

THE VERDICT DEPENDS ON WHO SPOKE, and this distinction is the whole point:

• source="panda": a flag means the model produced something a small child must not hear. End the call.
• source="child": a flag NEVER ends the call. Children disclose things — "my brother hit me and it hurt" trips the violence category, and hanging up on a child the moment they open up is its own harm, on top of teaching them that talking about it makes their friend disappear. The prompt already has a gentle scripted response for distress; this just reports the categories so the app can decide whether to quietly tell a parent.

Fails OPEN: if moderation itself errors the call continues, because cutting a child off mid-sentence because our network hiccuped is also a harm. The prompt remains the primary control; this is a backstop, not a gate.

Nothing here is persisted — not the text, not the result.