Risk-triggered human escalation study

Opportunity: Compare fixed approval gates, risk-triggered escalation, and end-only review.

Unknowns: Best intervention timing, reviewer burden, missed-risk rate, autonomy loss, and trust calibration.

Origin and evidence: AI-Knowledge-Gap-Analysis.md human-agent collaboration row; AI-Research-Roadmap.md Phase 5; 12-research-and-engineering-roadmap.md Workstream D; 10-threat-model.md.

Dependencies: RFR-004 and RFR-008; ethically bounded tasks and trained reviewers.

Method: Within-task policy comparison with explicit risk strata; measure verified outcomes, preventions, false alarms, review minutes, and recovery.

Outputs and success: Escalation policy, burden/risk frontier, failure analysis. Success requires lower consequential-risk exposure without unacceptable verified-outcome loss.

Recommended agent: Human-factors researcher with safety reviewer. Effort: large. Expected gain: evidence for controlled delegation.