AI
Read tools are not model authority
Tool execution is scoped by Otto context before a model answer can be grounded in school data. Write actions remain proposals until an authorized person clicks.
Every tool is read-only and exposes no write, send, approve, charge, enroll, refund, dismissal, or compliance-finalization action.
Tenant, user, membership, role, scopes, and permitted student ids come from authenticated request context, not model output.
Student reads require either a self scope or an authorized staff scope plus a student id in the permitted-student set.
Results carry source references, redactions, prompt-version keys, context identity, and human-approval metadata for later ledger joins.