Trust & Security

No black boxes. Every action on the record.

The approval trail is not a feature. It is the architecture. Every action has a named owner, a timestamp, and the information that was available at the time of the decision. The trail is permanent, searchable, and yours.

The quality gate at every output boundary.

PRQA (Propose-Review-QA-Approve) is the standard of care enforced at every output boundary in the grid. Nothing leaves the system without passing through this gate.

Propose

An agent prepares and proposes an action: a quote, a notice, a reply, a purchase order. The proposal includes the evidence, the assumptions, and any risk flags.

Review

The proposal surfaces to the appropriate human via the Surface inbox. The human sees the full context: the proposal, the evidence, and the prior history.

QA

Before an action reaches the commit gate, it runs through the domain-appropriate quality checks: format, completeness, risk flags, and compliance with standards of care.

Approve

A named human commits. The commit is timestamped. The record is written. The action proceeds. No agent can shortcut this gate.

The record that never lies.

The approval trail is the permanent, searchable record of every decision made through the grid. It answers every audit question before the auditor asks it.

Who decided

Every commit is signed by the named human who approved it. Agents cannot commit. The trail shows which human, which role, and which authorization governed the decision.

When they decided

Every commit is timestamped at the moment of approval. The sequence of decisions is immutable and verifiable.

What information they had

The trail records the proposal, the evidence, and the QA result at the time of the commit. You can replay any decision exactly as it looked to the human who made it.

What changed afterward

If anything changes downstream from a decision, the trail links the change to the decision that authorized it. The chain of custody is complete.

Tripwires, not just locks.

The Vault does not rely on perimeter defense alone. Guardian is a set of active detection tripwires designed to fire inside the perimeter, where an attacker who cleared the wall is most likely to act.

Decoys that detect

Synthetic data and fake credentials scattered through the system. Accessing them is an immediate alert. Legitimate work never touches them.

Canary trips stop everything

When a canary fires, the grid halts the affected scope. Not a log entry. A hard stop, followed by a human review before anything resumes.

Full rotation, not partial

When a credential is suspected compromised, the response is full rotation of the affected scope, not partial rotation that leaves attack surface.

ArcGate denies as signal

ArcGate (the perimeter firewall) treats denies as signals, not just blocks. Patterns of denied requests surface to the trust dashboard for human review.

Every agent has papers.

The grid beneath the surface is not a swarm of anonymous processes. Every agent has an identity, an owner, a registry entry, and an explicit grant of what it may propose. Authority is chartered, never assumed. This is what makes the machinery safe to forget about.

Registry entry. Every agent is registered with its name, owner, authorized scope, and current status. You can see every agent that has ever acted in your grid.
Scoped authorization. An agent authorized to draft quotes cannot file legal notices. Each authorization is explicit, narrow, and cannot be escalated by the agent itself.
Human owner accountability. Every agent has a named human owner. If an agent behaves unexpectedly, the accountability chain is immediate and clear.