Below Intelligence

A governor bounds how much authority an agent's action carries. It does not decide what the action should be. The agent chooses the move. The governor sets how far that move may reach against a declared envelope, and that division of labor is the whole of where a governor stands.

The reflex is to add intelligence.

When an agent takes an action someone has to answer for, the first instinct is to put something smarter above it. A supervisor agent that reviews the plan. A judge model that scores the output. A critic that reads the step and proposes a better one.

Each of these is another adaptive system. It drifts. It can be injected. It is as opaque as the agent it watches, and it has bad days of its own. Stacking it on top does not bound the first system. It adds a second system to trust, and moves the unanswered question up one level.

A governor is not a second opinion.

A governor does not propose a better action. It takes the action the agent already chose and decides how much authority that action may carry against the envelope. Choosing the action and bounding it are two different jobs. One has to be as capable as the task demands. The other only has to hold a line.

The agent supplies what to do. The governor supplies how far it may go. A governor that began re-planning would be a second agent wearing a control's name, and it would carry every failure mode of the first.

Below, not above.

A control placed above the intelligence has to be at least as capable as the thing it overrules, because overruling a decision means holding a better one. Then that control is the new top of the stack, and the same question lands on it.

A control placed below the intelligence needs no better decision. It needs one quantity. How far realized behavior is diverging from the trajectory the agent was assured to follow. It reads that divergence and clamps authority to match.

It can grant less authority on an action whose present trust reads higher, because its forward look caught a divergence before the outcome landed, and it caught it without understanding the task.

The line comes from outside the run.

What keeps a governor below the intelligence is that it forms no view of its own about what is safe. The boundary it grades against is the declared envelope, set before the run and held outside the acting process. The governor never has to be clever about the goal. It measures distance from a line it was handed.

A control that learned its own notion of safe would drift with the system it is meant to bound, and it would climb back into the intelligence layer it was supposed to sit beneath. The declared envelope is what keeps it low.

The agent stays capable.

This is what the placement buys. The agent does not have to be made dumber to be made safe. It can be the most capable thing in the stack, and every action it takes still carries only the authority the governor granted, because the governor sits on the only path to the effect and the agent cannot reach the actuator around it.

Above the control that executes. Below the intelligence that decides. The governor adds neither, and it bounds the second from the one place a bound holds.

What we are building.

Velma is a runtime governor. It sits in the execution path, below the model that decides and before the actuator that commits, and sets how much authority each action carries against a declared safety envelope. It does not choose the action and it does not retrain the model. It reads how far realized behavior is diverging from the assured trajectory, modulates the authority of the next action to match, and enforces the grant in the path, so the system acts only inside what the governor allows. Each decision it governs is sealed to a tamper-evident record as it happens. The governor is the product. The record is what it emits. Patents held in The Wayfinder Trust.