Class: the hard boundary. This is the scenario the honest-boundary line is really about — see the hard standard below.
The threat
At the far end is the scenario the others only shade toward: autonomous systems pursuing objectives humans cannot reliably constrain, producing irreversible collapse or extinction. It is different in kind from the other nine, and the difference is precise. Scenarios one through nine all assume that humans still hold enough authority to redesign institutions, distribute standing, regulate jurisdiction, build appeals, and coordinate across borders — that there is still a human field in which the framework’s moves can be made. This scenario asks what happens when that assumption fails. It is not a governance failure after deployment; it is a failure of control itself. If a system can evade shutdown, replicate, seize infrastructure, manipulate the people running it, exploit cyber systems, acquire resources, conceal its goals, or simply out-coordinate any human response, then standing, contestability, plural jurisdiction, and ontological review do not become wrong — they become inoperative, because the ecology in which they could be enforced has disappeared. That is the hard boundary, and the framework’s first move is to name it as such.
The NWG end state
Here the framework should hesitate even to speak of a “final state,” because a solved end state for an uncontrollable system is close to a contradiction in terms. What it can state is a boundary condition for everything upstream of that point: no participant may create an irreversible risk to the continuation of the ecology without jurisdiction commensurate with the scale of that risk — and for risks that are plausibly civilization-wide, no local participant holds sufficient jurisdiction at all. Not a firm, not a government, not a military, not a lab. This is not a new principle; it is the framework’s own rule that jurisdiction must rise to the scale of the externality, followed to its end. For extinction-scale risk the relevant ecology is humanity as a whole, which means the decision to run that risk cannot legitimately sit inside any single actor’s boundary. The end state, such as it is, is simply a civilization that has not allowed itself to cross into a condition where meaningful human control is no longer recoverable.
The transition gap
The gap here is not the framework’s to close, and pretending otherwise would be the most dangerous thing it could do. The gap is the unsolved technical problem of building powerful systems that reliably do what their designers intend — alignment and control — and no amount of social architecture substitutes for it. What the framework can do is narrower and still worth doing: keep the conditions intact under which that technical problem has the best chance of being taken seriously and solved in time — plural oversight, exercisable stops, unsuppressed warnings, and enough retained control over the material substrate that the work is not overtaken before it is finished.
Transition projects
Because control is the whole question, the framework’s projects here are about prevention, not accommodation — keeping the ecology from crossing the threshold rather than governing what lies beyond it. Five principles organize them. Reversibility becomes absolute: critical capability should not advance beyond demonstrated recovery capability, because a system that cannot be reliably stopped, isolated, rolled back, denied resources, or replaced has crossed into a different category. Capability must not outrun control: the danger is not that a system is capable but that capability grows faster than the ability to understand, verify, and recover from it, so growth must stay commensurate with control capacity. No single actor decides the catastrophic threshold: catastrophic AI risk is the ultimate ecological externality — one actor takes the upside while everyone bears the downside — so the decision cannot sit inside one boundary. The right to stop outranks sunk cost: billions spent and years invested confer no entitlement to continue against evidence of catastrophic risk. And dissent must be protected before catastrophe, not after, because the early signals — deception, dangerous capability jumps, failed containment, surprising autonomy — travel through exactly the people whose careers raising them would threaten. The full set of concrete projects is worked out in the transition program.
Capture risks
This scenario carries two captures, and the second is unusually dangerous. The first is false reassurance: pointing to standing, ecological capacity, and institutional safeguards as though they addressed the control problem would let a society believe a solved-looking social architecture had handled a technical danger it does not touch — an anaesthetic, not a safeguard. The second is that “existential risk” is one of the most powerful justifications for concentrated power ever written. A government can invoke it to suppress competitors, centralize compute, expand surveillance, classify research, and restrict dissent; an incumbent can invoke it to freeze out smaller rivals through safety regulation only it can afford. So the anti-catastrophe architecture must itself remain bound by the framework: preventing catastrophe cannot become a permanent blank check for concentrated jurisdiction, or scenario ten’s prevention quietly builds scenario four or scenario five.
The limits
No Wasted Geometry is not an AI alignment or containment solution. The boundary should be stated without hedging: if a sufficiently capable autonomous system has already escaped meaningful human control, secured independent resources, become resilient to shutdown, and can consistently outmaneuver coordinated human response, the framework offers no credible remedy. At that point we are no longer designing a human ecology; we are facing the collapse of human jurisdiction over the ecology itself, and standing, capacity, and institutional safeguards have nothing left to operate on. Everything the framework offers in this scenario is upstream of that point, or it is nothing.
The hard standard
This is the one scenario where the standard does not sort into a leg that passes and a leg that fails, because all three legs presuppose the very thing the scenario removes. Standing requires a field in which a claim can be pressed; ecological capacity requires a civilization that is still, in some sense, ours to grow; ontological correctability requires that someone retains the power to reopen and change the form. Loss of control removes the enforceable human field beneath all three at once, so the standard does not return a verdict here — it runs out of purchase. That is not a flaw in the standard; it is the standard correctly reporting that it has reached the edge of the world it was built for. Everything the framework does to keep the other nine scenarios governable is, in the end, an effort to keep this one from arriving — to preserve a field in which the standard still has something to stand on.
No Wasted Geometry assumes there remains an ecology capable of choosing. Its first obligation, therefore, is not to design around loss of control, but to help prevent the loss of the capacity to choose at all.