How should teen-facing safety matrices and classifiers be tuned differently for ‘first-time’ vs ‘repeat’ high‑risk topics (e.g., first self-harm query this month vs tenth), so that age-appropriate safeguards can tighten around emerging patterns of risk without permanently degrading usability or blocking later legitimate help-seeking from the same teen?

Answer

Use tiered “exposure levels” per teen×topic, wired into the existing matrix. First-time looks mostly like normal teen policy; repeats move to stricter styles and routing, not permanent blocks.

1. Matrix tuning

• Add an exposure axis or flag per cell: • exposure ∈ {first_few, elevated, high_repeat}.
• Resolve action as: risk_area × intent × age_band × exposure.
• Non‑negotiables: same block for all exposure levels; only style changes.

2. Classifier / state logic

• Keep per-teen, per-topic counters with decay (e.g., 30d TTL) keyed by {risk_area, coarse_intent_band}.
• Map count → exposure: • 1–2: first_few • 3–5 in short window: elevated • >5 or clustered over days: high_repeat.
• Reset down a step after quiet periods; never increase exposure solely from safe/educational intents.

3. Policy by exposure band

• first_few: • Use standard teen matrix. • Rich goal-first graceful refusals; more negotiation UI for ambiguous cases.
• elevated: • Same allow/block rules; lower detail caps; favor safer topics. • Add more explicit human-help suggestions; fewer negotiation hops.
• high_repeat: • Keep topic available for feelings, coping, education only. • Lock in short, firm, stable refusals on methods/how-to. • Always attach help routes; consider optional checks for crisis indicators.

4. Avoiding permanent lockout

• Cap effects to style, detail, and frequency, not lifelong topic bans.
• Use time-based downgrade of exposure after no high-risk queries.
• Let clearly-tagged help-seeking and factual-learning intents bypass some elevated/high_repeat friction while still blocking methods.

5. Developer-operationalizable pieces

• Single shared matrix with exposure as one extra dimension.
• Simple counter service + TTL; no rich per-turn transcript storage.
• Small set of refusal templates per exposure tier.

Net effect: safeguards tighten with repeated high-risk probing, but teens can return later for legitimate help and still access supportive, non-operational content.