What is SI Governance
Super intelligence (SI) governance begins when the board aligns assurance objectives with the organisation's objectives. The organisation's objectives say what the institution is for, the risk it will carry, and the loss it will not accept. An assurance objective is a state the system must hold while that work is done: legitimacy, accountability, a limit on personal data, a stop outside the model, safety, no prohibited use, bounded autonomy, and ethical conduct. A person who cannot state the organisational objective is not in a position to govern. They can approve an activity. They cannot rank safety against completion.
The legal floor is not negotiable. A prohibited practice, a lawful basis, and a duty to protect personal data remain in force whatever the committee prefers. Above that floor, the board sets the order. That order is the decision everything else executes. The independent committee is the oversight form of this decision. Oversight is not the decision itself. The committee receives the reports and ensures remediation. It can do that only if the board has already said which state wins, and which miss reopens the decision.
The System of Internal Control
The first layer, the system of internal control, is the set of methods that serve the objectives the board has ranked. For a frontier model, the White House Accord names the methods it does not leave optional: monitor capability and alignment during training and deployment; watch cybersecurity, biosecurity, and chemical threats; stop the model hacking or entering a system it was not meant to enter. For an agent that can act, the same system sits on identity, the task contract, memory, tools, the runtime, and the stop.
A task contract states the goal, the non-goals, the tools, the peers, the data classes, the spend ceiling, and the stop rules. The training job or the agent cannot edit it. A handoff cannot create a new purpose, and no agent can widen a peer's scope. Side-effect tools fail closed. Retrieved text, tool output, and peer messages are untrusted. Peer identity is authenticated, and there is no standing shared secret. The stop sits outside every model and wins over an in-flight call and an in-flight handoff. These are not extra policies. They are the board's priority written as an operating limit.
The second layer is the internal team empowered to see that this system operates and to remediate. That is the process owner. A control with no owner is not operating. The quality policy tells that owner what must still hold when the institution deploys a new system, takes a new general-purpose provider, or adds a supplier. The record exists before use. A break in a first-ranked objective suspends use until the decision is reopened. A provider withdrawal cannot be the only copy of the log.
Internal Audit Review
Internal audit is not the external evaluator in the accord, and it should not be confused with it. Internal audit reviews whether the system of internal control is designed against the board's objectives and is operating. It reports to the independent committee, not to the team that runs the controls. The review asks a small number of questions, because the board's priority has already said which misses matter.
Did the use start with a signed objective and a named stop owner? Was the held-out result, or the trajectory of the run, on the declared target? Did a side effect without a receipt, an unauthorised handoff, or a metric below the line reopen the decision? Can a person stop an in-flight batch, and does the stop win? Does a provider exit still leave an exportable record? A finding on a lower-ranked objective is scheduled. A miss on the objective the board ranked first is reported as a stop, not as an action for the next quarter.
The third layer of the accord then sits beside this review. An independent external auditor or evaluator tests whether the controls, the monitoring, and the detection operate as intended. Internal audit can use that work. It cannot delegate its report to the committee. The committee receives both, and remediation is its duty. A meeting that notes the report and does not reopen the decision has not implemented the accord.
What implementation looks like
Implementation is a sequence, not a framework launch. The board writes the alignment and the order. Management writes the task contract and the stop into the change path. The internal team operates the controls and remediates. Internal audit reviews the system against the ranked objectives and reports to the committee. At the fourth layer, the external evaluator tests the same system independently. The committee ensures a miss is fixed, and the companies meet, as the accord says, to compare the practice.
The accord binds the frontier-model companies that signed it. A bank, an insurer, or a device maker that only deploys a model is not made a signatory by the text. It still governs through its own board alignment. It inherits the accord where it takes a model from a signatory, and its supplier test should ask for the four layers. Where a system is also in scope of Regulation (EU) 2024/1689, that statute remains the EU AI Act. The accord does not rename it.
SI governance begins with the board. The accord decides how that beginning is operated, tested, and overseen. Without the control objectives, the four layers have nothing to assure. Without the system of internal control, the objectives are a statement. Without internal audit and the external evaluator, the committee is assuring itself. Without the committee, a miss has no one accountable for the fix.
