In-path
Policy before weights.
Governance & Audit
Policy, DLP, human-in-the-loop and audit live in the path of inference. Logs are written where the workload runs. A report generated after the fact is not governance.

In-path
Policy before weights.
HITL
Consequential actions wait.
Immutable
Logs where the work runs.
Offline
Controls survive a pulled cable.
The Gateway presents the request. Policy matches data class and caller.
Allow, redact, deny, or hold for a named approver.
Inference or tool call proceeds only on an allow. AgentLab waits on HITL for consequential steps.
The decision, model version, caller and outcome are written to an immutable log on the same machine.
Which model class may see which data class. Named rules, not tribal knowledge.
Fields that must not leave a document unmodified. Redaction is logged. Originals are preserved.
Consequential AgentLab actions wait for a named approver. Nothing silent posts to ERP.
Written locally. Not a cloud SIEM you hope is reachable.
Who may invoke, who may approve, who may read the log.
Pull the network. Policy still matches. Logs still write. Approvals still wait.
The rule fires before the weight sees the prompt.
Modify is an application action. The audit of it is a Foundation duty.
AgentLab does not execute silent writes to systems of record.
The security pack lists the fields. The install writes them.
A weekly export from a SaaS console is a report. It is not a control.
If governance requires a vendor cloud, it will fail the day the cable is pulled.
DocxIntel asks and AgentLab actions share the same policy engine.
Licence list, audit fields and key custody are files. Not a slide.
In your office, on your documents, with the cable pulled out. No competitor's sales engineer can do this.