The model sees minimized references — not raw data.

Context engineering decides what the model is allowed to see and act on. Host context is projected down to references and counts before the model runs, stripping raw and restricted content, so the model receives the minimum it needs to reason — not the underlying records.

Boundaries are enforced by the platform, not by prompt discipline. Cross-entity and cross-conversation data mixing is hard-blocked before the model runs, returning a context-boundary violation, so one customer's data can never be pulled into another customer's session. Entity resolution works by reference and never fetches raw records directly — raw content is only ever read through governed tools, so reads obey the same permission and audit path as writes.

The platform also watches for drift. It baselines the underlying data and re-probes at run and resume; when the ground truth shifts, stale context becomes a soft blocker instead of the agent acting on out-of-date assumptions. And every result, detail, and comparison output is labelled with what it was based on, so users can judge and verify an answer rather than relying on it blindly.