The Anchor
and the
Forkable World

01 What the Diagnostic Actually Says

Part II established that durable work requires someone who bears irreversible loss for getting a decision wrong, who is recognized by the community the decision affects, and whose consequences can’t be insured, indemnified or replayed.

That diagnostic survives. What didn’t survive the framework’s own stress test was the promise hiding underneath it: that this durable layer was large enough to absorb the displaced knowledge workforce, if people would only organize their lives around it. That promise was wrong. The durable layer is real, but it is very small, and it is small for a structural reason, not a temporary failure of imagination.

The diagnostic still does what it was built to do, with full force: it helps you see what you’re looking at. What it can’t do is turn clear sight into a living for everyone who looks honestly.

None of what follows is a moral claim about who deserves security, dignity or political protection. The subject is structural leverage in an economy increasingly run by machines, not moral desert. Keep those two categories strictly separate as you read.

02 Forkability and Fragility

The clearest way to see why the durable layer is small is to ask what makes a consequence impossible to fake.

An AI agent is forkable. You can copy its weights, snapshot its state, shut down the running instance and start a successor with the new lesson built in. The system loses essentially nothing. Its “consequence” amounts to an update. Whatever moral weight you assign to punishing it, the substrate itself is cheap to replace, and the agent is reborn or swapped out at trivial cost.

A human being is fragile, in a specific technical sense: one instance, which can’t be rebooted, backed up or restored from a saved state. Death destroys everything that instance was. A collapsed reputation can’t be patched. Prison spends a finite resource, time, that can’t be reissued. Public disgrace doesn’t roll back to the morning.

The point isn’t that a human role can’t be refilled. Roles are refilled every day. The point is that the particular person who bore the loss can’t be restored. A successor surgeon, judge, parent or founder can step into the role, but the person who was lost stays lost. The role survives; the bearer doesn’t.

Mortality works here the way a private key works in cryptography. The signature can’t be forged, because the thing that produces it can’t be copied.

As far as anyone can tell today, human beings are the only available agents that combine unrecoverable loss with things a society can actually hold accountable: recognizable personhood, social standing, and the whole repertoire of human consequence, from punishment and disgrace to forgiveness, mercy, trust and exile. That combination is what makes humans so hard to substitute in certain roles.

It also explains the shape of the terrain. Wherever a system requires an agent whose loss can’t be recovered by copying, and whose loss the community can see, humans are the only option. Wherever a system will accept a forkable agent whose “loss” is just a software update, the economy will eventually route the function through that cheaper agent.

Institutions do this routing through the wrapping mechanism from Part II. They place a fragile human behind a forkable decision-maker, collect the appearance of consequence from the human, and route the actual decision through the AI, which can absorb any amount of “punishment” at no cost to itself.

So the diagnostic question for any role becomes:

Do the person’s fragility and the authority to decide sit with the same person? Or is a fragile person bearing the consequences of a forkable agent’s decisions?

If they sit together, the role is bedrock. If the person is bearing consequences for the machine, the role is wrapped work in formal dress, however prestigious the title.

The Tesla Autopilot litigation shows how the split plays out, and also that it can be contested. In 2025 a Miami jury held Tesla 33 percent responsible for a fatal crash involving Autopilot and assigned most of the fault to the driver. The driver was fragile and bore most of the blame. The Autopilot software was forkable: Tesla paid money, and any lesson learned could reach the whole fleet in an update. The case isn’t a pure scapegoat story, since the driver was not paying attention and had real control of the car. But the asymmetry is exactly the one this section describes. And the verdict, which survived Tesla’s attempt to overturn it in 2026, shows that courts can push consequence back upstream onto the forkable side.

The corollary is uncomfortable. Many of the “human-in-the-loop” requirements now being written into AI deployment preserve the form of human judgment without the substance. They are scapegoat positions engineered in advance: the safety driver, the person who signs off on the AI’s medical recommendation, the approver of the algorithmic loan denial. They supply accountability theater while the forkable system makes the real decision. Learn to recognize them. Refuse them where you can.

03 The Fusion Test

A role is bedrock only when one and the same person holds all five of these properties. Fusion is the word for it: authority and consequence welded to a single fragile bearer.

If any one property is missing, the role is wrapped.

Apply the test to your own work. Most readers will find that the parts they assumed were bedrock are wrapped, and some of the parts they assumed were marginal carry the real weight.

04 The Bedrock

The bedrock is the set of roles where fragility, authority, control, consequence and traceability stay fused on one person. They cluster into five recognizable shapes.

This is the bedrock. It is real, and it has structural protection.

But it doesn’t employ eight billion people. Counted as a headcount, it is a small fraction of today’s knowledge workers: on my judgment, a few percent at most even if strong institutions are built to protect it, and well under one percent if they aren’t. Treat those figures as orders of magnitude, not measurements.

If that were the whole story, the rest of this essay would be a eulogy. It isn’t the whole story.

05 A Small Bedrock Is Not Small Demand

The first draft of this essay made a quiet mistake. It counted how many people could stand on the bedrock and treated that number as the size of the human future. But a headcount of bedrock roles and the share of the economy that flows to human beings are two different quantities. Confusing them makes the picture darker than it needs to be.

Economists have a name for the mechanism I missed: Baumol’s cost disease. When productivity explodes in one part of the economy and stays flat in another, the flat part’s goods and services get relatively more expensive. If people can’t easily do without them, those sectors take a growing share of total spending. That’s why, as farming and manufacturing got radically more productive over the twentieth century, health care, education and personal services absorbed more and more of what people spent. In 2017, the economists Philippe Aghion, Benjamin Jones and Charles Jones applied the same logic to AI. If some tasks resist automation and are needed alongside the automated ones, those tasks become the bottleneck, and their share of the economy can rise rather than fall.

Applied to this trilogy, the argument runs like this. An AI economy could grow much faster than the human population. Even if only a modest fraction of it requires human beings, a modest fraction of something growing that fast can be a great deal of work, and better paid than today.

Everything turns on one condition: whether any human work is genuinely hard to substitute. The cautionary tale is the horse. Horses had comparative advantage too, but once engines could do everything a horse did more cheaply, the “wage” for horse labor fell below the cost of feeding the horse. America’s horse population peaked in the 1910s and then collapsed. A booming economy didn’t save them, because nothing they offered was hard to replace.

This is where the forkability principle earns its keep. Sections 02 and 03 are, in effect, an argument that one category of human work can’t be substituted: work whose value depends on a fragile, accountable, recognizable person bearing the consequence. Add the work where being human is the product itself: provenance, presence, care, performance, witness. If that category holds, the Baumol logic runs in humanity’s favor. The cheaper everything else gets, the more these become worth.

It also changes how we should read the protected-provenance category from Section 04. Artisans and masters don’t mainly need protection from weak demand. As machine-made goods get cheap, the willingness to pay for verified human origin is likely to rise. What they need protection from is forgery. France’s appellation system is useful because it certifies origin and punishes counterfeits, while the market supplies the premium. That is a narrower and more achievable job for institutions than propping up a craft against its own customers.

Three things can still make this go badly, and they move the bleakness rather than remove it.

Who holds the purchasing power. Demand for human work needs money behind it, held by someone. If the gains from AI concentrate among the owners of capital, demand for humans may still be large, but it will look like a patronage economy: many people employed by a wealthy few as artisans, performers, attendants and companions. Pre-industrial elites employed enormous numbers of servants. That is employment, but not the kind of stakes-bearing life this trilogy describes.

The ceremonial seats may be the ones that survive. Regulation that requires human oversight of AI is spreading, and legal demand for someone to hold accountable could keep many wrapped seats alive. The number of jobs might hold up while the judgment inside them drains away: the scapegoat positions from Section 02, multiplied. I think this is the most likely way the optimistic version fails.

Timing. Even good transitions front-load the pain. During Britain’s industrial revolution, output rose for decades while working-class wages stagnated, a stretch the economic historian Robert Allen called “Engels’ pause.” Gains that arrive later than the losses still wreck the generation caught in between.

So the real question isn’t whether an AI economy will want human beings. On this argument it probably will, perhaps more than ever. The real question is who holds the money, and whether the human work it buys carries real stakes or only their ceremonial shell. That is a political question, not an engineering one, and it is still open.

06 The Strategic Reserve

Beneath the question of livelihood sits a more basic one: the case for human existence in the AI epoch at all. That case rests on what the literature on existential risk calls preserving options under deep uncertainty about what ultimately matters.

The folk version: you don’t delete a codebase roughly four billion years in the making just because you don’t need it on Tuesday.

Strategic Reserve / living floor

A system that is genuinely uncertain about what is ultimately valuable can’t rationally destroy a unique repository of variation and adaptive capacity that can’t be rebuilt. The expected cost is too high. And a frozen archive isn’t the same thing as a living reserve. The value of keeping options open comes from evolutionary and cultural processes that continue, which requires living, dying agents that can’t be rebooted.

The argument is sound, but its conditions are cold. It binds only a system with a long time horizon, one that values keeping options open and is aligned enough not to discount human options to zero. It defends against one specific kind of panic about displacement. It is no guarantee against a catastrophic deployment failure or a fight over resources.

Within those conditions, it says humanity continues as a recognized presence because deleting it would destroy options irreversibly. It doesn’t say the AI economy will pay you; Section 05 is about that. It says the species continues, provided whatever is allocating resources is aligned enough to recognize the argument.

That isn’t the consoling version. It is the structural reason the lights stay on.

07 Track One: The Apex

If you’re in the bedrock, your situation is specific. Your role is durable because of the forkability principle, not because any institution guarantees it. Nothing yet stops an institution from quietly wrapping your role the moment it becomes convenient.

If the right institutions get built, the apex stabilizes with real authority and leverage. If they don’t, it shrinks to a small elite whose personal prestige is the only thing holding them in place. The apex needs four kinds of institutional support:

Records alone just produce cleaner stamps for wrapped work. They matter only when paired with real consequence.

None of these institutions exists yet in the form the framework requires. The nearest thing in law is the European Union’s AI Act, which obliges organizations deploying high-risk AI to assign human oversight to named people. That provision could create bedrock seats or scapegoat seats; the fusion test is how you’d tell which. Building the real thing is a project measured in decades, and the apex will either take it on or lose its footing.

If you pass all five properties of the fusion test, your work is clear. Defend your role institutionally, and train successors. Apply the diagnostic in public. Defend the boundary between bedrock and wrapping in the open record. And understand that your durability is a contingent fact you are responsible for preserving. An apex that fails to build its own housing has written its own extinction.

08 Track Two: Everyone Else

This is where the framework owes the most direct honesty. Most of today’s knowledge workers are outside the bedrock and won’t get in. If that’s you, or you’re being pushed toward it, this section refuses to lie to you in three specific ways.

• It won’t tell you that retreating into bedrock work is a viable economic strategy for most people. The forkability principle explains why the bedrock is small, and the institutions that might widen it don’t exist yet.
• It won’t tell you that you can escape the imitation economy through personal virtue. Synthetic experience is going to be the everyday medium of the AI epoch, and you will live inside it. The question isn’t how to exit. It’s how to take part without mistaking simulated satisfaction for a substantial life.
• It won’t tell you that a tidy political solution will arrive in time. Past waves of technological displacement were slow, and they routinely produced decades of hardship before adaptation caught up.

What the framework can offer are tools for living well in a bounded position, and, after Section 05, one economic opening worth taking seriously.

Honest recognition isn’t escape. For some people, accepting a bounded position will be a relief, the end of an exhausting effort to optimize against an economy whose terms have permanently changed. For others it will be unbearable. The framework’s job isn’t to resolve that grief. It is to give you the honest geography.

09 The Honest Summary

The bedrock is real and small, and it is protected by the plain fact that a human life can’t be forked. The Strategic Reserve is the civilizational floor beneath everything else, and it holds only on conditions. The apex can turn the bedrock into economic leverage, but only if it builds the institutions to protect it. And the much larger human economy outside the bedrock depends on a question this trilogy started out ignoring: who holds the purchasing power.

Part I said value moves to the boundaries. Part II said most of those boundaries are wrapped. Part III says the unwrapped layer is very small, the principle protecting it is robust, and the institutions needed to turn that protection into leverage are unbuilt. It also says that a small bedrock doesn’t doom the rest of us, because an economy where machine work is cheap may value human work more, not less, if the gains are shared.

That gives the future two independent questions, not one. Will the institutional housing get built? And will the gains from AI be broadly held, or concentrated among those who own the machines? Cross them and you get four futures.

This isn’t a prediction. It’s the decision surface. All four futures are live, and which one we get depends on choices made over the next few decades, most of them political.

Coda

This trilogy isn’t a survival manual. It is a terrain map, drawn by people who took the displacement seriously enough to look at it without flinching.

We ran the structural arguments against the strongest adversaries we could find, and we’re reporting the geography honestly. In places it is smaller and colder than we wanted. In one important place it is less bleak than we first drew it.

What the trilogy refuses to do is turn these tools into a false promise of economic leverage.

The balloon was not the point. The view was not the point. The wrapping must not be allowed to swallow the unwrapped layer without an active institutional defense. The Strategic Reserve holds when nothing else does, and it holds only on conditions. The work that doesn’t compress is the work that runs on lives that can’t be backed up.

The AI epoch is not the apocalypse, and it is not a renaissance. It is a redistribution of leverage along an axis the previous era didn’t measure. It will look like injustice to some and vindication to others. This framework’s job was not to argue with that geography, but to map it accurately enough that you can cross it from wherever you stand, carrying whatever you have, toward whatever bounded meanings you’re willing to commit to.

The life-work continues either way. The paid work continues where purchasing power and institutions route value back to the people doing it. Use what is useful. Discard what is not. Build with the recovered capacity rather than against it.

Whether you take part in that work is your choice. The framework has no authority to make it for you.