The governance layer behind every agent we run — what gets checked, how it fails closed, where the audit trail lives.
Checks run before the draft is queued.
The policy stack runs at generation time, not after. Five checks, configurable per agent — fit to the job, not a generic filter.
Voice match
Output reads as your voice. Tuned to your tone, register, vocabulary.
Claim verification
Any factual claim is checked against the linked source before the draft is queued.
Policy compliance
Per-agent rules — read, write, trigger — enforced before generation.
Confidentiality
Sensitive fields stripped before the model. PII never sent to the model.
Brand consistency
Approved terminology and prohibited language enforced across every output.
Pass, warn, or block — configurable per agent.
Every check returns one of three outcomes. Pass lets the draft move on. Warn surfaces a note in the queue — the draft still goes out. Block stops the draft entirely.
Severity is configurable per agent and per policy. A voice mismatch might warn on one agent and block on another. You decide what is a hard rule and what is a soft one.
If the policy layer is unavailable, the agent does not run.
The policy layer is not optional. If it cannot be reached — network failure, outage, schema error — every agent fails closed. Drafts are not produced. Actions are not taken.
Fail-closed, not fail-open. A missed draft is recoverable. A draft that bypassed policy is not.
Policies are data, not code. Versioned and reversible.
Every rule lives in the policy store — versioned, scoped, auditable. Org-wide applies to every agent. Agent-specific overrides at the agent level. Changes are admin-managed and logged.
Roll back in one click. Diff any version against its predecessor. Export the full set for compliance review.
Two synced logs. Reconstructable months later.
Two synced logs — one internal, one exportable. Every input, output, and decision is recorded with timestamp, actor, and context. Append-only.
When you need to answer a compliance question — what did this agent do, when, and why? — the answer is in the trail. Filterable by agent, actor, action, time.
Not a substitute for human approval.
Not consumer content moderation. Not a marketing safety net. Does not catch every edge case. Does not replace the judgment of your team.
Agents operate under an approval gate. Every output is a draft until a human approves it. That is the contract.
We share the policy schema, audit-log export format, and fail-closed behavior on request. If your security team needs to review before an engagement, we'll walk through it with them.