The Golden 30 Minutes: Controlled Dialogue Design of an AI Collaborator for Crisis Intervention in Special Education

In-class crises among mainstreamed students with special needs (emotional outbursts, self-harm tendencies, teacher–student standoffs) demand intervention within the golden 30 minutes, yet general-education teachers often mishandle events due to a lack of validated scripts. General-purpose conversational AIs further risk cultural mismatch and overreaching recommendations. We propose G30F, a controllable AI collaborator framework with three operationalized layers: (i) a content-safety layer combining a multicultural lexicon (218 high-risk terms, R1–R5) with emotion-intensity thresholding (E1–E7); (ii) a structural-control layer encoding authoritative crisis guidelines into a three-stage finite-state machine (Soothe→Guide→Resolve/Refer) with explicit circuit-breakers; and (iii) an ethics-constraint layer enforcing human authority via a Teacher–Policy Weight (TPW) with automatic fallback to observer mode. We instantiate G30F on a 7–8B open-source backbone trained via SFT(LoRA) → DPO → PPO with a composite reward. On 1,840 teacher-labeled events from 118 schools and a two-phase human study of 200 dialogues, G30F improves stage-match accuracy by 19 pp, cultural safety by 12.3 pp, and ethics compliance by 14 pp over single-objective baselines (p < 0.05, Holm–Bonferroni), while remaining deployable on teachers' existing devices.