Judgment stays human.

The claim.

Judgment is human-owned, visible, and refinable.

Three words, each load-bearing.

• Human-owned — a person stays accountable. Not "a person was in the chain somewhere." Accountable.
• Visible — the judgment can be inspected, in language a human can actually read. (What "visible" means in practice →)
• Refinable — when the judgment is wrong, the correction lands on the judgment, not on a downstream symptom.

If any of these three slips, you don't have human governance. You have a logbook.

What AI actually does inside this frame.

AI assists. It refines. It searches. It routes. It executes under constraint. These are real and substantial — none of them are governance.

The distinction matters because executing is the fastest of the five. When execution speed outruns the human's ability to inspect what was decided, the system starts deciding without you. Quietly. At scale. Often correctly enough that nobody notices the seam — until it matters.

What cannot be silently delegated.

• The choice of what's worth doing.
• The judgment of whether the answer makes sense.
• The decision to act on it.
• The acknowledgment of responsibility for the outcome.

These can be assisted. They cannot be handed over and then called "AI governance" with a straight face.

Approval is not judgment.

A signed-off PR is not a code review. A signed audit log is not an audit. An "approved" button is not a decision — it's a button.

The trap: build a system where the human's only role is to approve what the machine prepared, and you've created the surface of judgment without the substance. Reviews become rubber stamps. Sign-offs become defenses-against-blame. The judgment that was supposed to stay human migrated quietly into the system's defaults, where nobody is watching it.

We design against this on purpose.

The four control surfaces.

The architecture under this page names four surfaces where judgment shows up — the places governance has to land if it's going to land at all:

• Intent — what are we aiming at, and what aren't we?
• Judgment — how do we decide what's good, bad, risky?
• Causality — how do results actually arise, and how does learning compound?
• Relational — where can clarity land in this system, and where can't it?

Most "AI governance" frames address one of these — usually a slice of causality (logs) or implementation (permissions). The Verse view is that all four have to stay legible at once, or the system can convince itself things are fine while the four are quietly drifting in different directions.

What this means for the systems you build.

If you're shipping AI inside an organization right now, the question worth asking is not "do we have governance" — most teams have a governance page. The question is: can a human still see and own the judgment in the loop, in language they can read, fast enough to matter?

If you want a structured way to find out, the Signal Profile is where that conversation starts.