The scenario describes a progressive, step-by-step reduction of AI autonomy as risk increases. This is a defining feature of the Graduated Response governance pattern within the CAIPM framework.

Graduated Response is designed for high-risk environments where a binary on/off control (such as a kill switch) is insufficient. Instead, the system dynamically adjusts the level of AI control based on real-time conditions. In this case, the system begins with minor restrictions (limiting valve adjustments), escalates to requiring human consent for each action, and ultimately removes the AI entirely if the situation remains unstable. This tiered escalation ensures safety while maintaining operational flexibility.

Other options are less precise:

Boundary Constraints impose fixed limits but do not evolve dynamically with risk escalation.

Kill Switch represents an immediate, complete shutdown rather than a phased reduction.

Disengage Capability refers to the ability to remove AI from the system, but not the gradual escalation process described.

CAIPM emphasizes that in safety-critical systems, graduated control mechanisms allow organizations to balance efficiency and safety by scaling AI autonomy up or down depending on risk conditions.

Therefore, the correct answer is Graduated Response , as it best captures the systematic, risk-based reduction of AI agency.