A school-district agent must handle student PII and curriculum-policy text carefully. Which context design fits?

Separating student PII context from curriculum-policy context deliberately is the correct design for school-district agents because privacy boundaries and retrieval boundaries should align. Student identifiers, grades, and contact data belong in tightly controlled case context with strict access logging; curriculum standards and pacing guides are broader policy material that can be retrieved without attaching a child’s identity. Deliberate separation reduces accidental leakage into logs, prompts, and shared indexes while keeping policy guidance available when needed.

Storing student PII with curriculum policy using type tags for filtering fails because type tags are only a logical filter, not a boundary; both kinds of data still share the same retrieval surface, so a query error or tag bypass can leak a child’s identifiers into policy results. Embedding student identifiers in curriculum-policy vectors to personalize retrieval fails conceptually—it bakes protected education data directly into the reusable policy corpus, so any policy search can surface or expose a specific student’s identity. Redacting student PII at presentation while keeping identifiers in retrieval context fails because the identifiers remain in the model’s context window and retrieval logs, where they can be exfiltrated or woven into outputs before any presentation-layer redaction happens.

Exam caveat: some tasks legitimately need both (IEP-aligned tutoring); join them ephemerally under least-privilege session controls rather than storing student PII with curriculum policy using type tags for filtering. Operational check: trace a tutoring session’s retrieval calls, confirm PII is never embedded into the policy index, confirm policy queries return no student records, and red-team a prompt that tries to exfiltrate roster fields via vector search.