The scenarios name what can go wrong. The irreversibility register names the lines a civilization must not cross. Graduated jurisdiction and ownership under review say that consequential power must stay bounded. All of it describes the order that should hold — and none of it names the agent who imposes it. That is the hardest gap in the framework, and the one a serious critic reaches for first: a design is not a seizure of power, and the transition’s worst danger — the concentration of AI capability and of the ownership behind it — is exactly the actor with every incentive to refuse the constraint. This page takes that objection at full strength. It is about who actually constrains concentrated power, why that is harder than getting the powerful to agree, and how the institution built to do the constraining keeps from becoming the next concentration.
Constraint is leverage, not belief
The mistake to avoid at the outset is imagining that constraint follows from the powerful accepting that they should be constrained. It does not. Durable constraint comes from the ecology controlling something the powerful actor needs and cannot cheaply route around — and there are many such dependencies: compute, energy, capital, specialized talent, market access, physical infrastructure, legal recognition, insurance, and legitimacy itself. Constraint becomes credible exactly when losing access to one of those would cost the actor more than compliance does.
So the operative question is never “who believes in this framework?” but “who holds a dependency this actor cannot escape?” This reframes enforcement as a map of who depends on whom, not a hierarchy of who has formal authority over whom — and it is why unbundling ownership matters so much in practice: the powers bundled inside large-scale ownership are also the levers through which its exercise can be conditioned. A constraint with no such lever behind it is a warning, not a limit. That is the whole content of enforcement capacity: the ecology’s real ability to hold a line against an actor with reasons to cross it.
The enforcement concentration paradox
Here the framework meets a problem it cannot dissolve, only manage. To constrain an actor operating at the frontier of capability, the constrainer may need comparable reach — privileged access to the systems, deep technical competence, the compute and instrumentation to inspect them, perhaps systems of similar power. The institution strong enough to check a dangerous concentration therefore risks becoming a dangerous concentration itself. Stated plainly:
The stronger the capability being constrained, the greater the enforcement capacity required; beyond some threshold, the capacity needed to constrain a concentration may itself become a concentration no one can undo.
This is the enforcement concentration paradox, and it is not answered by adding the words “oversight” or “appeal.” Those help at the margin; the structural trap remains. A framework built to keep any function from becoming the whole system has to apply that suspicion to the enforcer above all — because the enforcer is the one function explicitly granted power over the others.
Prevent the concentration before it needs a constrainer
The paradox has a direction, and the framework should follow it. If the institution strong enough to cure a frontier concentration is itself a concentration risk, then the cleaner path is to keep the concentration from forming in the first place — so that no frontier-scale constrainer is ever required. That promotes the framework’s forward-looking tools from options to the primary line of defense: graduated jurisdiction’s insistence on deconcentrating what is existentially dangerous, ownership under review’s limits on how much jurisdiction may gather in a single place, and pacing’s refusal of an entitlement to speed. Constraint after the fact is the fallback for what slips through, not the plan.
Prevention is preferable to cure not because cure is unavailable, but because the cure for a sufficiently large concentration is itself one.
This is the mature form of the framework’s dual doctrine of power: build the capacity to constrain concentrated power, and constrain the capacity that does the constraining. And there is a reason prevention cannot simply be deferred until a concentration becomes intolerable — the means of preventing it do not stay available.
When constraint must be built anyway, bound the constrainer
For concentrations that already exist, enforcement has to be built, and then the same doctrines that bound every other authority apply to it — turned upward. The enforcer is an authority like any other, so minimum necessary jurisdiction, expiring authority, and the rule that duty of care scales with power all bind it, inside the constitutional layer that ordinary majorities may not cross. Concretely, that means an enforcement architecture assembled from several bodies rather than one: inspection separated from sanction and from operational capability; multiple independent technical evaluators rather than a single authoritative one; no single body holding all the relevant data; powers that are revocable and sunset by default; overlapping jurisdictions that can check each other; hard limits on reusing inspection data, so that verification never quietly becomes surveillance; external review; and technical capability that is federated across many hands rather than monopolized in one. The enforcer, in other words, is itself entered on the irreversibility register and watched like any other concentration — the framework never stops doing power analysis merely because an institution has “public” written on the door.
Even so, one question stays open even after every safeguard is in place: can enforcement kept deliberately distributed and correctable actually keep pace with a concentration that is none of those things? The framework does not get to pretend it knows.
Leverage has a half-life
The last difficulty is that even the right leverage decays. A dependency that constrains an actor today can be routed around tomorrow as that actor internalizes its own compute, energy, capital, and talent — and rising AI capability is, among other things, a general engine for dissolving bottlenecks. So enforcement readiness is not the static question “do we hold a chokepoint?” but the dynamic one “how fast is this chokepoint becoming routable?” That rate is a leverage half-life, and it turns the real test of a transition into a race: is the ecology building durable constraint faster than the actor is dissolving the leverage that constraint depends on? A regulatory system that looks strong on paper may already be near impotence if the dependencies beneath it are quietly disappearing.
This exposes a red line the register did not previously name — one that sits upstream of the loss of human control:
Governance becomes non-traversable when the actors being governed can route around every dependency through which public constraint operates.
At that point the systems may still answer to their operators, and technical control may not yet have failed — but the public has lost any practical hold on the operators, and course-correction is gone before the catastrophe the loss-of-control scenario describes ever arrives. Governance non-traversability is the governance failure that precedes the technical one, and it is exactly the kind of decaying condition the register’s pre-irreversibility emphasis exists to catch while turning back is still possible.
The balance is old; the speed is new
None of this is unprecedented, and the framework should say so rather than pretend to have invented it. The idea that liberty survives only where a capable society and capable public institutions hold each other in balance, and that concentrated power is checked not by a single benevolent authority but by many distributed centers of countervailing power, is one of the oldest lessons in political economy. What the transition adds is not the balance but the rate. Traditional balance assumes that firms, states, and civil society all evolve at roughly human organizational speed. Rate asymmetry breaks that assumption: if one center can compound its capability far faster than the institutions meant to check it can perceive, organize, legislate, and build a response, then the forms of plurality can remain standing while the substance quietly disappears. The measurable version of the danger is a countervailing latency — the time opposition needs to notice, coordinate, and act, set against the actor’s own doubling time; when the second runs much shorter than the first, effective plurality is already gone. This is captured as countervailing latency, and it is where the framework’s contribution to an old problem actually lies.
It also forces one concession. A norm that cannot attach to a material lever does not constrain power, however well-grounded the norm is — legitimacy, ideas, and institutions become consequential only through structures able to convert them into action. That is a real concession to the materialist reading of power, and it must be scoped precisely so it does not undo the rest of the framework: materialism governs whether a norm constrains power, not whether the norm holds. The non-subordination that grounds standing is untouched by any of this. “Materially enforceable” is the test of a norm’s political effect, never of its validity — the participant does not stop counting because the ecology currently lacks the leverage to protect them.
What this settles, and what it does not
The framework, then, has a strong account of why power must remain contestable and a growing account of how constraint can be made credible: through leverage rather than goodwill, through prevention before cure, through an enforcer bound by the same doctrines it enforces, and through a clear-eyed view of leverage that decays. What it does not yet have — what no one has — is an answer to whether enforcement that stays distributed and correctable can keep pace with frontier systems whose capability and infrastructural independence may grow faster than any institution constraining them. That is the real open problem, and hiding it would be its own kind of failure.
The problem is not only how to build enough power to constrain power. It is how to do so before the leverage to do it disappears — and without letting the constraint itself become the next concentration no one can undo.