Part I established that human override is not an external safeguard. It operates inside the system, as a delegation surface through which authority expands when it is not explicitly constrained. Correction at the level of the instance becomes, over time, a mechanism by which authority grows.
The institution that discovers this reaches for the obvious fix: make the system legible. Instrument decisions, surface deviations, trace patterns across time and risk class. Monitoring frameworks, evaluation pipelines, audit trails, and reporting layers go in, and they quietly take on a second role. They stop being instruments of visibility and start standing in for governance.
They cannot. Visibility does not constrain authority. A system can be fully observable, legible at every layer, and still operate within the same set of permitted actions. Observed that thoroughly, monitoring does not constrain the system. It describes it.
Visibility Without Constraint
Observability is not the problem. Enterprise AI deployments track model performance across many dimensions, evaluate outputs against curated datasets, log and surface decision pathways, and record, categorize, and trend exceptions. Governance forums receive ever more detailed pictures of system behavior.
All of it answers one question: what is happening? None of it answers a different one: what is the system still permitted to do? Visibility surfaces patterns, deviations, and trends. It does not decide which of those behaviors remain permissible. A system can be precisely observed, every decision recorded and every exception traced, and operate within the exact set of permitted actions it had before anyone was watching.
That is enough to manufacture the appearance of governance. Behavior is no longer opaque, trends are discussed with precision, the system is interrogated through data instead of anecdote. The surface reads as controlled. The permitted action set has not changed. What looks like control is observation without constraint.
The Illusion of Oversight
Visibility substitutes for control most easily where control is expensive. Complex authority surfaces, distributed across workflows, embedded in dependencies, coupled to throughput, are costly to change: changing them slows execution, disrupts process, and forces coordination across the organization.
Observation costs none of that. It layers onto existing systems and scales without asking them to change. Over time the cheap thing stands in for the expensive one, and the ability to watch a system in detail passes for the ability to govern it. Evaluation practice reinforces the swap: models are continuously assessed, outputs measured against known distributions, deviations cataloged, the system steadily more legible to the people responsible for it. What does not grow is the mechanism that turns that legibility into a change in what the system may do.
The patterns are all visible. Override clustering in a risk class can be seen. Escalation paths compressing under load can be detected. Applied posture diverging from approved posture can be surfaced. None of that is constraint. An institution can watch a boundary come under strain and leave it unchanged, watch exceptions become routine and keep processing them, detect authority expanding and have no way to contract it. Seen in full, governed not at all: that is oversight without control.
From Detection to Constraint
The question is not whether drift can be observed. It is where, and how, authority is changed.
In most enterprises those are different places. First-line operators (1LOD) execute within defined boundaries and see the system directly: where decisions cluster, where overrides accumulate, where escalation paths compress under pressure. Second-line functions (2LOD) turn observation into posture: they define risk classes, set thresholds, and own the governance artifacts that determine what is permissible. Between them, change is mediated on purpose. Observed patterns become proposed updates; classifications are revisited, thresholds adjusted, policy rewritten, reviewed, approved, and only then implemented back into the system.
That mediation is slower than observation by design, and the velocity gap is where the system keeps operating with the authority it already has. Closing the gap is a design problem, not a discipline problem. Watching more carefully does not shorten it. Four things do.
Pre-define the constraint for each monitored signal. For every signal that matters, override clustering in a class, escalation latency compressing, applied posture diverging from approved, name in advance the contraction it triggers: which reachable actions narrow, which class suspends, which path starts requiring escalation. The signal and its response are specified together or not at all.
Pre-authorize standing contractions at the first line. The slow path exists to change policy, and it should stay slow. A bounded set of protective contractions can be delegated to 1LOD in advance, to fire the moment a signal trips, ahead of formal review. Narrowing authority is safe to make fast; widening it is what must stay deliberate. Let the reversible, restrictive move happen immediately and reconcile it through 2LOD after.
Make monitoring emit constraints, not just reports. Route the signal into the enforcement path directly, so a tripped threshold reduces the reachable action set rather than generating a finding for a forum to discuss weeks later. Observation that cannot reach the authority surface will never govern it.
Measure the see-to-change latency itself. The interval between detecting drift and contracting authority is a governance metric in its own right. Track it per risk class. When it grows, the institution is accumulating exactly the exposure this series is about: a system observed in full and constrained not at all.
Closing
Monitoring is a means, not an end. Bind every signal that matters to a change in what the system may do; a signal that cannot reach the authority surface is instrumentation. The test is one question: when monitoring shows drift, can the institution contract authority in response? If it cannot, the system is observed, not controlled.