The library returns a verdict and decides nothing about what follows. Something has to sit where claims cross and act on what it returns, and that something has to be introducible into a plant that will not be rebuilt to accommodate it. This section fixes where the enforcement point sits, what it is, and what a plant does with each verdict.
Where it sits
The review already located it. On the Purdue model the segmentation perimeter guards the crossing from the information-technology levels into the plant, and admissibility sits downstream of that, inside the operational-technology zone, immediately before the act. The perimeter asks whether a value may cross; admissibility asks whether the act it would justify is warranted, and those are different questions asked in different places.
That location has a consequence for what the enforcement point can be. It cannot live in the transport, because transport does not know what act is coming: the same value crossing the same conduit may inform a display, prompt an investigation, or authorise a valve movement, and only the last of those has a demand worth enforcing. The demand belongs to the act, so the enforcement point belongs where the act is decided, which is at the consumer.
It is not one thing, and only some of it is enforcement
A plant has more than one place where a claim is turned into an act, and they are not the same component. There is the point of actuation, where a controller acts on a value. There is the supervisory boundary, where a decision is made about what to command. And there is the human interface, where a value is presented to an operator who will act on it.
Each takes the same form, because the library's interface is the same everywhere: a record, a demand, a verdict. What differs is what the point can do with a refusal, and the difference is not incidental. Refusing at an actuator blocks a physical action. Refusing at a supervisory boundary blocks a command. Refusing at a human interface blocks nothing at all, because a person can act on what they have been shown. What happens there is disclosure rather than enforcement: the claim is displayed with the fact that its basis did not meet what this act requires, and whether that changes anything depends on the reader.
The distinction is worth keeping sharp, because the first two make an act unavailable and the third informs a decision that remains available. Only the first two are enforcement in any strict sense. Calling disclosure by the same name would suggest the discipline had gated something when it had only reported.
Disclosure is the weakest of the three and it is worth having anyway, because it is the case the review documented most heavily. The human-factors literature is about people acting on conveyed state whose basis they could not appraise, and a display that shows the verdict alongside the value gives them what the discipline can supply and what those accidents lacked.
How it is introduced
The enforcement point wraps rather than replaces. It sits between an existing consumer and the value it already receives, calls the library, and passes the value on or does not. Nothing about the transport changes, no protocol is added, and the systems carrying values continue to carry them.
This is a constraint the implementation section already fixed for a plain reason: an artefact requiring a plant to adopt new transport to gain admissibility would have made the discipline conditional on a change nobody will make. Wrapping is what makes it introducible one consumer at a time, which is also how it will be adopted if it is adopted at all, since no plant will place a new gate in front of everything at once.
What a plant does with each verdict
The response is configured per act and recorded, not fixed by the library. Three verdicts and three decisions, and the reason for configuring rather than fixing is that a discipline which blocked on every refusal would be removed.
Admit passes the value on and records that it did. The record matters: an admitted claim whose act later proved wrong is the case that tests whether the demand was adequate, and that is not recoverable if only refusals are logged.
Refuse is configured between blocking and alarming. Blocking is right where the act is consequential and a wrong act is worse than no act. Alarming is right where not acting is itself dangerous, which the review established is common in operational technology, and where the value of the discipline is that a person is told the basis was inadequate rather than that the act is prevented. Both record the verdict and the failing axis.
Undetermined is the configuration that decides whether the discipline is being run or performed. Treating it as refusal is conservative and right for consequential acts. Treating it as admit is legitimate where the missing axis was never load-bearing for that act, and dangerous everywhere else, because a plant with incomplete records and a permissive undetermined setting passes almost everything while appearing to run the discipline. The artefact cannot prevent that configuration; what it can do is require it to be stated per act rather than defaulted globally, so that the choice is visible to anyone reading the configuration.
What placement does not solve
An enforcement point can be bypassed. A consumer can be rewired to receive the value directly, and in a plant under operational pressure it will be, if the gate refuses often enough to stop useful work. That is the failure the methodology names as over-conservatism, arriving as a physical fact rather than a measurement, and no placement prevents it.
What placement can do is make a bypass a change rather than an omission. Because the enforcement point wraps a consumer rather than being embedded in it, removing it is an alteration to the plant's configuration, which is recorded and reviewable in the way a plant already reviews such alterations. That is a weaker guarantee than the discipline being unavoidable, and it is the honest one.
One assumption in this section should be marked rather than left implicit. The argument for disclosure at a human interface is that a person integrates a refusal with everything else they know, which is why showing them the verdict is worth doing even though it gates nothing. That reasoning depends on the recipient being a person. Where a probabilistic system occupies the reviewing position, a refusal placed before it is a text it may weight, ignore or reason around, and the case for disclosure as a substitute for enforcement does not carry across unchanged. The conclusion returns to this.