Zero-Click Guardrails for Classroom AI: A Policy-Compiled Ethics Kernel with Pre-Execution Adjudication

In contemporary classrooms, AI systems do more than generate content; they broker instructional decisions-task selection, difficulty calibration, timing, reframing, and feedback routing. Along this decision surface, latent risks-prompt manipulation, privacy leakage, plagiarism laundering, and policy circumvention-often evade ad-hoc safeguards. We propose a zero-click classroom AI guardrail: no model output is enacted unless admitted by an ethics-verification kernel. The kernel encodes institutional policies in a domain DSL, maps them to temporal-logical constraints, and performs real-time pre-execution adjudication with auditable traces. The architecture comprises three tiers: (i) policy and signal repository, (ii) verification plus continuous auditing, and (iii) decision orchestration with adaptive caching. We evaluate the kernel in four schools, three subjects, and ten weeks of deployment. Metrics include capture rate for unsafe recommendations, false rejections, latency budget, and teacher-in-the-loop acceptance. Results show protection against prompt-injection, identifiable information leakage, and jailbreak attempts, while maintaining instructional flow within acceptable delay bounds. We distill principles for multilingual use, cultural alignment, and compliance with local regulations. Finally, we outline practical implementation guidelines for LMS integration, policy exchange, and evaluation artifacts. The proposed guardrail strengthens trustworthiness, transparency, and harm reduction in AI-mediated instruction, using diagrams and audit logs rather than equations.

Hali tarjima qilinmagan