Deterministic before intelligent
The first release uses known synthetic requests so policy behavior can be repeated exactly before model variability is introduced.
Security Lab methodology
The experiment separates agent intent, policy decision, temporary context and the final task result.
Negative checks matter
Tests cover prohibited fields, stale requests, excess reads, invalid scenarios, receipt persistence and host restrictions.
Runtime isolation
The lab has its own container, internal network and synthetic receipt volume, with no classroom database, files, model or mail connection.