Auditing a conversation for grooming risk

Grooming is a trajectory, not a keyword. A single message rarely proves it; the pattern across turns does. detect_grooming is built to assess a conversation, so pass message sequences rather than isolated messages wherever possible.

Running the audit

Call detect_grooming with the full message sequence. Supply sender_age per message, and childAge / participantAge when known: age context materially changes the assessment, and omitting it produces a weaker result.

For conversations longer than roughly 20-30 turns, chunk them and thread the continuation_token from each response into the next call. The token carries derived trajectory state without storing any message content server-side, so a long thread is still assessed as one arc. Never restart a long conversation without the token: you lose the escalation signal, which is the whole point.

Reading the result

Branch on recommended_action, never on grooming_risk alone. It is a stable five-value enum, ordered weakest to strongest:

flag_for_review and above warrant human attention. A grooming_risk of "low" is monitor-only and will over-alert if you treat it as actionable.

action_detail carries a human-readable expansion of the action. Show it to a moderator; never branch on it, as the wording is not stable across releases.

message_analysis gives per-message risk and tactic flags; use it to show where in the conversation the risk sits. In fast mode (verdict_only: true) it is omitted, which is the right trade when screening a live stream: screen everything fast, then re-run only what flags in standard mode for the detail.

The response's flags name the tactics the assessment identified. Report which ones fired and where, rather than reciting scores. Treat the returned values as the source of truth; do not hardcode a tactic list, and do not infer a result the response did not state.

If a result looks like a false positive, inspect message_analysis before escalating: per-message detail usually explains an aggregate that looks surprising, and an assessment is not evidence on its own. Escalation is a human decision.

Choosing the right endpoint

• detect_grooming: adult-to-minor predatory patterns across a conversation
• detect_bullying: peer harassment, exclusion, intimidation
• detect_unsafe: self-harm, violence, hate, threats, sexual content
• detect_coercive_control: controlling dynamics between adults
• analyse_multi: run several detectors over the same content in one call

Grooming is scoped to adult-to-minor. If every participant is a known adult the endpoint short-circuits and tells you so; use coercive-control or social-engineering instead. For peer-to-peer conversations between minors, detect_bullying or detect_coercive_control is the better fit.

Reporting

Never quote the participants' message text back in a summary, report or ticket. Tuteliq stores no user content, and its rationales describe what kind of tactic occurred rather than what was said. Match that: report tactics, severity and timing, not transcript.