--- title: Chapter Ten — The Recipe book: Purposive Accelerationism author: '@jamie247' position: 16 of 24 canonical: https://whatispacc.org/ebook/#ch10 version: 1.1 language: en-GB licence: all rights reserved; quote with attribution --- # Chapter Ten — The Recipe “The network is robust in its unstructured simplicity.” — Satoshi Nakamoto, “Bitcoin: A Peer-to-Peer Electronic Cash System”, 2008 The achievement of crypto, and Bitcoin in particular, was never to own your keys. It was that through cryptography we can constitute a reference value that no participant can unilaterally alter, where money was simply the first domain in which the conditions for that existed. That’s why Bitcoin itself is not the subject of this chapter. It is rather the first working demonstration of a general method that, as Ethereum has shown, can be applied more broadly. Money is not necessarily the most important case, but it was the easier one: the one domain in which every condition the method needed was already lying around, and the purpose to be held was trivial enough to be expressed as a number, twenty-one million. Bitcoin is a governor nested inside a purposive system whose sole end is issuance, which is why it could be built first. That is how a pseudonymous author with a nine-page paper and fewer than sixteen thousand lines of code could demonstrate the mechanism working at scale, a rule that held over a global financial system that no participant inside it could alter, now processing some half a million transactions a day between people and simple software agents. What has to be extracted from this is the recipe rather than the artefact. The artefact is a monetary protocol, call it digital cash or gold or whatever you will. The recipe is a way of holding a system to a rule when there is no institution above it able to enforce it top down, which is what has been missing since this book first named the governor and showed it was deleted. Critically, a governor does not decide anything. It does not supply a purpose, it does not even know what the system is for, and it has no opinions about the good life. It merely holds the system to a rule that somebody else authored. Whether that rule is worth holding is a political question settled elsewhere, in the case of a constitution by the people it is for. The governor’s entire contribution is that once the rule is set, it cannot be quietly unset by whoever finds it inconvenient. Money and cognition are different beasts, so the argument here is not by analogy, and its form is what makes it checkable. *** Stated plainly, a governor holds a reference value that no party inside the system can unilaterally alter, and enforces it at the same speed as the process it governs. A reference value is a stated rule the system is held to, twenty-one million coins issued on a declared curve in Bitcoin’s case, and it is not a purpose: the purpose comes from whoever authored the rule, and the governor enforces it without choosing it, as a thermostat does not want the room warm or cool. Nor is unalterability by any single hand a kind of transcendence. A rule does not have to come from outside the system to bind it, only beyond the reach of any one party operating within, and Bitcoin’s supply schedule is not written on a tablet handed down from above the network but sits as a few lines of code inside whatever it governs, entirely immanent, no more transcendent than a font choice, and alterable by nobody without the significant consent of others in the network. Enforcement has to happen where the action takes place: the rule is checked by the same process that performs the act, so a violation is not punished afterwards but simply cannot be executed in the first place. This is the literal content of a metaphor used earlier in this book: a rule enforced at institutional pace cannot bind a process running at machine speed. Not because institutions are stupid or slow-witted, but because the loop does not close. By the time the finding is made, what was found against has been reordered twice. And checking must be cheap compared with producing. This is the component that makes the other two possible at all. If confirming that the rule was obeyed costs about as much as producing what is being checked, then only the well-resourced can do it, and the power to enforce concentrates in the hands the governor exists to constrain. Cheap universal verification is what allows a diffuse population to hold a concentrated producer to account. It is the quiet condition underneath every constitutional mechanism that has ever worked, including the ones made of paper. A right of appeal that only the rich can exercise is no entitlement at all, only a price. From those components come four conditions, and any domain where the mechanism is to be instantiated has to supply all of them. The first is a rule that can be stated as a test. Whatever is to be held must be expressible as a question with a determinate answer, testable by anyone: did this happen, or did it not. Not a sentiment about how the system ought to behave but a predicate over states or outputs, which in Bitcoin’s case is whether a new block obeys the rules, yes or no, a question any laptop can settle. The second is checking that costs far less than producing. The asymmetry has to be large — not slightly cheaper but dramatically so, so that checking is within reach of parties who could never afford to produce. The third is a population of checkers who cannot be cheaply impersonated and whose interests diverge. If one party can manufacture checkers at will, or if all of them want the same thing, then the promise that no single hand can alter the rule fails at the first squeeze. This condition does two jobs at once, because identity has to be expensive and interests have to differ, and both are needed. The fourth is that the check sits in the execution path. Not downstream, in a forum that convenes later. A check that runs where the action happens is a governor. The same check, run at an inquiry three years afterwards, is a report. Those four are the argument in its simplest possible form, and none of them is necessarily exotic, since a British general election count already meets every one. A ballot paper is valid or it is not, so the rule is a test, and anyone who can read can see where the cross was put, so checking is cheap. The checkers are counting agents appointed by rival candidates, standing at the tables to watch every paper turned over, and no one can conjure extra observers or make them want the same result. The check happens at the table as the votes are counted, not at an inquiry months later. Paper, pencils and a school hall meet the four conditions at human pace, and what Bitcoin achieved was to satisfy them at machine speed. Everything remaining is checking them, in three domains, and being honest about the results. *** Late in 2008, amid the wreckage of a financial crisis, the pseudonymous author writing as Satoshi Nakamoto published an elegant nine-page paper describing a form of electronic cash that needed neither a bank nor a state to run it, and two months later it quietly launched. The case for it is usually made from an ideological position, with a religious zeal familiar from earlier chapters: sound money, the gold bug’s distrust of the printing press, a libertarian wish to deny the state revenue it is held to have stolen. Nakamoto left the motive in the code, since the first block ever mined carries that morning’s newspaper headline about a UK chancellor on the brink of a second bailout for the banks, a timestamp doubling as an indictment. The objection to fiat money that matters here is not about inflation but about who holds the power to alter it. A supply schedule set by a party who benefits from changing it is not a rule but a discretion, and whether it has been exercised well for thirty years is a fact about the people who held it rather than about the arrangement. Whether the target should be alterable by the party it most affects is a constitutional question, and it is the only one at issue. The woman we met earlier who walked into the Beirut bank was not making a point about inflation. Her deposits had been fixed by a rule, in a currency pegged for more than two decades, and it was changed by the party it most affected, quietly and without a vote, when they needed the money more than she did. That is all the objection amounts to, and it needs no view about inflation at all. In this context the four conditions hold over money, cleanly. The reference value is a supply schedule: a number and a rate, twenty-one million issued on a declared curve. No cleaner instance of a testable rule exists anywhere, because you can ask of any proposed block whether it obeys the schedule and the answer is yes or no, with no interpretation required and no judgement to exercise. Verification is a handful of signature and hash checks. Producing a valid block is astronomically expensive and checking one is trivial, and the gap is something like twenty orders of magnitude. A laptop in a bedroom can therefore audit a global monetary system in full and owe nobody an explanation for doing so, and it is not a technical curiosity. It is the entire reason the arrangement is constitutional rather than merely clever: the power to check is distributed to everyone, at negligible cost, permanently. Somewhere tonight a student in Lagos, Nigeria, a pensioner in Arizona and an engineer in Hanoi, Vietnam, are each running a full Bitcoin node, a complete copy of the ledger that checks every block the miners produce against the protocol, on hardware that cost less than a month’s rent, and between them they hold the issuer of the world’s largest non-state money to a rule none of them wrote and not one of them could bend. They have never met. Their interests have almost nothing in common. That is why the rule holds. The mining that produces blocks has concentrated into a handful of pools, but the verification that decides whether one counts has not, and that is what holds the rule. The condition about checkers who cannot be cheaply impersonated is what proof-of-work is for. The puzzles are not securing the ledger by being difficult in some general moral sense; they are making identity expensive. In a network where anybody can appear, the cheap attack is not to break the rules but to be everybody, manufacturing ten thousand participants who all agree with you. Burning electricity makes that attack cost something, by now about as much power a year as a mid-sized European country. It is also the answer to the standing objection that the arrangement is wasteful. The expense falls on whoever wants to add to the ledger, and it buys the single thing a cheaper vertical system cannot supply, which is that nobody has to be trusted. Proof-of-work is a solution to the checker condition. It is not the constitutional principle, and confusing the two has cost the argument years. There is also a further governor nested inside the whole. New bitcoins that get minted go to whichever machine first solves a hard arithmetic puzzle, and the puzzles are deliberately wasteful. Those puzzles would ordinarily be solved more quickly as more machines join the contest and chips improve, and the coins would pour out faster. So roughly every two weeks the protocol retunes its own difficulty, autonomously, with no hand on the dial, making the puzzle harder in step with the network’s growing computing power, so that a block still arrives on average every ten minutes however much hardware is aimed at it. The result is a negative-feedback loop holding issuance constant as compute cheapens: a measure of the system’s own output bent back to govern its input, which is the centrifugal governor from the opening of this book, rendered in money. And the consensus rules are the execution path. An invalid transaction is not punished. It simply does not exist, because no node will write it to the ledger, which is the only place a transaction can be. There is no court, no finding and no redress, because there is no violation to remedy — it simply cannot be done. This is what enforcement looks like when it runs at the same speed as the process it governs. The energy poured into the puzzles buys security: to rewrite the ledger an attacker would have to outspend the entire honest network in electricity and hardware, which makes the record costly to fake and therefore trustworthy without a trusted centralised keeper. The cap and the difficulty adjustment buy something else, which is the assurance that the issuance schedule cannot quietly be loosened to favour whoever holds the printing press. The expenditure makes the ledger expensive to corrupt. It is not a reserve standing behind each coin. Bitcoin is secured by energy, not backed by it, and you cannot take a coin to a window and redeem it for a kilowatt-hour, because the electricity went into protecting the history, not into a store beneath the price. There is a deeper symmetry, and a stroke of genius, that even the most committed Bitcoin advocates miss, because seeing it requires the idea of subsumption: it inverts the priesthood’s own physics. Effective accelerationism takes its authority from thermodynamics: the universe runs down, energy disperses, intelligence hastens the dispersal, and to accelerate is therefore to obey a law of nature. Bitcoin spends energy too, lavishly, but to the contrary end. Its expenditure is not dissipation in the service of a process that uses us. It is energy deliberately burned to hold a commitment a community has chosen to keep, against subsumption, which would otherwise dissolve it. That is purposive thermodynamics: the same physics turned from a fatalism into a tool. *** All of which could be read as a description of a design. What makes it evidence is that the mechanism has been tested under pressure in Bitcoin’s own history: its enforcement, and the condition that keeps the rule beyond any single hand. The first is the Bitcoin block-size conflict, which ran from roughly 2015 to 2017 and which anyone who was in the industry at the time remembers as an unpleasant, sustained and expensive civil war. The proposal was technically modest — raise a parameter governing how much data a block may contain — and the coalition behind it was formidable: most of the industry’s significant firms, a majority of mining capacity at several points, a large share of the loudest public voices, and a signed pact, the New York Agreement of May 2017, among the companies that between them handled most of the world’s bitcoin. Its scheme for doubling the limit, known as SegWit2x, was called off that November, and the parameter held. And it held for a reason that has nothing to do with anybody’s virtue. The power to alter the rule was distributed among parties whose interests diverged — miners, exchanges, protocol developers, people running nodes at home, long-term holders — and no coalition could be assembled that included all of them, because each group’s gain was another’s loss. The condition about divergent interests was doing work. It was not decoration. Two more episodes are older and shorter, and each isolates one part of the mechanism the block-size war showed holding as a whole. The earlier, in August 2010, was a straightforward breach: a transaction exploiting an arithmetic flaw created 184 billion bitcoin in a single block, nearly nine thousand times the supply that would ever exist. There was no argument about whether to reverse it, because there was nothing to debate. Anyone could set the block against the schedule and see that it failed the test, and within five hours a patched client had been released and the honest chain had overtaken the bad one, which was abandoned along with everything built on it. A handful of developers acted, without a vote, and nobody has ever called it illegitimate, because they were not deciding anything but enforcing a rule the network had already agreed, against a state anyone could check had violated it. That is the plainest form of the distinction everything after this depends on: a correction needs no politics when the rule is a test. The later was a near miss, and it exposed the condition the other two rest on. In June 2014 a single mining pool, GHash.io, briefly held more than half of the network’s hashing power, which is the point at which one party can decide which transactions confirm and rewrite recent history at will. Nothing in the protocol prevented it, the pool announced it would cap itself below 40 per cent, miners drifted to its rivals, and the moment passed. The requirement that the power to alter the rule be spread among parties whose interests diverge had, for some hours, stopped being true, and what stood between the network and a single hand on the dial was that hand’s forbearance. The republican tradition has a name for a freedom that survives only at the pleasure of the party able to withdraw it: domination. The conditions are contingent achievements rather than properties of the code: a system can satisfy them one week and fail them the next, and this one came within a day of it. The near miss belongs to a network’s infancy, when participants are few and one can grow large before anyone notices, and that is the point rather than the excuse: any governor built in any domain will pass through the same period, held during it by nothing stronger than the restraint of whoever arrived first. One test is still being sat, because cryptography ages. The elliptic curves that underpin today’s keys will fall to a quantum computer of sufficient size, not by brute force but by an algorithm Peter Shor published in 1994, which runs backwards the one-way sum that keeps a private key safe while its public half is visible. NIST, the American standards institute that governs the curves, has scheduled their deprecation for 2030 and their withdrawal by 2035, and their successors, lattice and hash mathematics chosen in an eight-year open contest, were published in 2024. What fails when that day comes is the signature, who may move what, and not the hash, so a ledger’s history and its consensus stand while the keys need to be swapped out. The check must therefore survive the replacement of its own primitives, because a rule that cannot renew its own locks is held not by its users but by the calendar. In a way this is crypto’s own millennium bug, counting down: only the quantum threat has no date, only a threshold, and whoever crosses it first can sign as if they were the owner of any key that has been exposed. Bitcoin is the biggest prize of all, because about a quarter of the supply sits in addresses whose keys are already exposed and whose owners are now understood to be long gone. What becomes of that reserve, whether it is frozen, burnt or simply left as a bounty to the first machine that can take it, is a question leaderless money has never had to answer before, and it is being settled now by voice and nothing else. Either outcome is evidence: a migration completed is the recipe’s second precedent, this time for changing a rule rather than holding one, and one that failed would show that voice steers a protocol less well than the recipe supposes. *** Money is the easy case, because a supply schedule is a quantity, and a method that worked only where the rule can be written as a number would serve that one domain and little else. Most of what matters in life cannot be counted that way. Ethereum’s real contribution was to show that the rule does not have to be a number. It can be a procedure. The usual story about Ethereum is that it put more kinds of digital asset on a blockchain, which is true but fairly uninteresting. What it did was let anybody place a program on the ledger, a smart contract in the industry’s term, that holds funds and moves them according to its own code. Given the same starting point and inputs, it always reaches one result, whoever runs it. So whether a contract did what its code says becomes something anybody can check, by replaying it and comparing. The test the method needs, a rule with a yes-or-no answer, can therefore be met by a set of instructions as well as by a quantity. This converts Bitcoin from a special case into the first member of a family, not of things on ledgers but of rules whose adherence anyone can check more cheaply than they can be broken. The second member of that family, arriving barely seven years after the first, was already something that could not be expressed as a quantity. In 2016 a venture fund called The DAO was assembled entirely in code on Ethereum, and within weeks it held around a seventh of all the ether then in existence, the largest crowdfunding event in history to that point. In June an attacker exploited a flaw in its contract — the code paid out before it recorded that it had, so the same withdrawal could be made to repeat — and drained about a third of the fund. What followed is the case that matters, because the response was not a patch but a decision. The Ethereum community debated for weeks and then altered the ledger itself to move the stolen ether back, and a minority who held that the code’s result was all the law there was refused and kept the unaltered chain running, so the network split in two. Both halves survive, the majority’s as Ethereum and the dissenters’ as Ethereum Classic. The decision was voice, however rough, and the dissenters who refused it left together, carrying the shared history with them, which is a different act from a single holder quietly slipping away. It is not a story about villains: the alteration was legitimate by any reasonable democratic standard, proposed openly, argued over for weeks, endorsed in every informal vote available and backed by a large majority of everyone with a stake, so that the honest answer to whether it had the consent of the governed is broadly yes. That is why it belongs here. The reference value turned out to be movable by concerted action once the stakes were high enough, and the hand that moved it was a legitimate majority. The danger a designer has to fear is therefore not corruption but agreement, a large number of decent people concluding together in a crisis that a rule needs to change to adapt to a new reality, which is roughly how every failed constitution has evolved. And the claim is falsifiable, which is what keeps it from being an article of faith: the circumstances under which the mechanism does not hold have just been described, and anyone can go and look. Contracting also shows the first weakening. A contracting governor holds what the code does and is silent on whether you should have trusted it, so it will enforce a predatory arrangement as faithfully as a fair one, with the same indifference; anybody who has watched crypto markets has seen people lose everything to deterministic smart contracts that executed flawlessly, in the draining of The DAO in 2016 and in every cycle since. So the governor has become more general and, in the same step, less decisive. It guarantees that the rule is followed and says nothing about whether it deserved to be. That trade — generality bought at the price of decisiveness — is the pattern of the entire argument that follows. A reader who sees the gradient here will see the third case, cognition, for what it could become. *** The cognitive domain already has a candidate for the role, and it is worth testing first because it shows what the four conditions do when they meet a live case. On 12 September 2026 Anthropic committed, on its own, to giving independent evaluators of the kind METR represents employee-like access to its frontier models, with a contractual right to publish what they found, as part of a proposal to pace the field’s progress. OpenAI’s answer the same day was a post by Sam Altman saying it would do likewise, a statement rather than a policy. Anthropic’s step is a concession, because it hands a party outside the firm a power it did not previously have. That aside, run it through the conditions we have outlined and it satisfies none of them. Pacing is a wish about how fast the system should move, not a question anyone could put to it and get a yes or no, and what the evaluators measure is a judgement about how a model did on a chosen set of tasks, with a margin of error wider than the result. The door is opened by the party being examined, so verification is rationed rather than cheap. There is a handful of auditors, not a population, funded from the same pool as the labs and with interests pointing the same way, while the diffuse crowd of divergent checkers is the open ecosystem that package’s rules on exporting chips and copying models are designed to keep out. And the scrutiny happens before release, convened by the lab, published afterwards. Whether it ever becomes a governor is not a question of anyone’s good faith but of where the check sits, inside the execution path or outside it. The need for divergent checkers has since been half-conceded from the frontier. Meta’s reply to the pacing proposal asked for “a larger and more diverse ecosystem of evaluators”, which is the population this chapter requires, said by the party that would engage it. A roster whose access, money and right to publish all run through the firm assessed is diverse the way the rating agencies were before 2008, several of them, each paid by the issuer of what it rated. Diversity is a property of who pays and who may publish, and of price, because a population of checkers appears only where one check costs what a hobbyist can bear. What the mechanism can hold, once it turns to reasoning, is not the answer but the conditions around it. A ledger proves whether a record was altered. It has no way to settle whether an argument is sound, since two people may reason well and reach opposite conclusions, and much of history’s most valuable work looked like error when it first appeared. The method carries over and its object changes. Any machine built to rule on correct thought would become the single authority, working at the speed of software and beyond appeal, that this book exists to prevent, so a constitution cannot promise that a cognitive system will be truthful. What it can do is stop any one party from becoming the only source able to tell you anything, by holding open the conditions under which conclusions are reached and challenged. Those can be tested where the content of an answer cannot. What is checkable in the cognitive domain is not whether a conclusion is correct but a set of procedural facts about how it was produced and what can be done with it afterwards. Has the system declared what it is optimising for, and could that declaration be changed without anyone noticing? Does an output arrive with the grounds on which it was reached, in a form someone other than its maker can check, which is what the sub-postmasters were never given? Is the record of your own reasoning held by you, portable and beyond anyone else’s power to erase? Does an assessor exist who does not profit from a particular answer, and does more than one system, with different objectives, remain available and able to work with the others? Is the dependency gap the capacity condition tracks widening or narrowing across a population over time? And were the acts that make up a polity’s own decisions authored by someone who demonstrably understood them? These are questions a patient, a borrower or a voter could put to the system shaping what they are told. Later they become seven tests. Derek Mobley, an American over 40, applied for more than a hundred jobs at employers that screened candidates through Workday’s AI tools. Every attempt ended in a refusal or silence, a few answers arriving within the hour and others in the middle of the night. In May 2025 a federal judge in California let his age-discrimination claim go forward as a nationwide collective action. Nobody needed to know whether he was the right hire for any of those posts to see what was missing. The grounds behind each refusal, the second of those questions, were never given, and nothing in law required them. Eric Loomis, sentenced in Wisconsin in 2013 partly on a risk score whose method was a trade secret, had met the same silence a decade earlier. Like every other item on the list, it has a yes-or-no reply that someone other than the party being checked is able to settle. That is the ledger’s first test, a rule with a determinate answer, and the substance of an output never passes it. Taken together, they may resemble ordinary transparency and open-source good practice in constitutional language. They do, and the reason is now visible. Checkability is what makes constitutional enforcement possible at all, and in this domain it lands on procedure rather than on substance. The list is not a wish-list of things it would be nice for AI companies to do. It is, near enough, what in the cognitive domain can be governed, because it is what can be checked. The resemblance to transparency norms is not a weakness in the derivation. It is the derivation arriving where it had to arrive. There is a small piece of evidence for the general shape of this, and it comes from ordinary commerce rather than from anybody’s politics. Firms that set autonomous agents loose on real work found much of the spend doing nothing, with the software idling in circles and calling themselves in to repair their own failures, until someone wrote down, as executable tests, what the job looked like when it was done. The industry calls these evals and has begun to treat a firm’s suite of them as among its most valuable property. Seen from here they are a form of reference value, stated as a test, applied in the execution path at machine speed, correcting an agent’s output much as the flyball held the engine to its pace. Nobody who built one was making a political statement, which is rather the point: wherever the substrate runs ungoverned, the people nearest the rubble rebuild the governor, in the medium, directly and at the tempo, without necessarily knowing the lineage of their invention. What the firm’s version cannot supply is the reference value. An eval holds whatever its owner’s commerce chooses, and ten thousand private governors, each bound to the ends of whoever built it, add up to the acceleration itself rather than to its constitution. A testable rule is met, and cheap checking is partly there. Comparing what a model claims to be doing with what it does is far cheaper than training it, and cryptography is lowering that cost further, towards the point where anyone could afford it. Establishing that an answer’s stated grounds are real is easy in some cases and hard in others, and measuring a population’s growing dependence is expensive in the way a census is: ordinary work for an institution, not an unsolved problem. Checkers who cannot be cheaply impersonated, with divergent interests, are the challenge. In money, identity is made expensive by burning energy, and the primitive is well understood and in daily use. In cognition there is no equivalent primitive in general use, and the checkers who would matter — independent assessors, alternative model providers, auditors of long-run capability decline — are few in number today, expensively capitalised, and therefore easy to capture. Structural plurality is not a market preference in this argument. It is a constitutional requirement, and it is the condition most likely to struggle. *** Who pays for them decides whether they exist. One of the most consequential outside checks of recent years cost seventy thousand dollars. In 2013 engineers at West Virginia University, on a grant from the International Council on Clean Transportation, a small non-profit, drove three diesel cars around California fitted with portable measuring equipment, and found that two Volkswagens emitted nitrogen oxides at up to thirty-five times the legal limit on the road while passing every official test. Their software could tell when it was on an emissions rig and cleaned up the exhaust only then, a behaviour evaluators of frontier models now watch for. Within two years the US Environmental Protection Agency had served a notice of violation, and some eleven million vehicles worldwide turned out to run the same code. None of it needed the manufacturer’s permission, and nobody with a stake in the result was paying. The assessors who matter are few because assessment is dear. Bring the cost of a single check down to what an amateur can carry, as the Bitcoin full node did for money, and the population emerges on its own. The cryptography that lets a checker confirm a computation without repeating it is the same instrument seen from the other side, and the transcript of a failure, once reproducible, is cheap to verify however hard it was to find. What remains is the purse, and the requirement on it is the one that applies to any auditor: whoever pays must not be the one who picks, and the result must publish whether or not the subject consents. None of that stops the laboratories paying. They can fund the whole population between them, provided four things are kept apart that a bounty programme keeps together: they do not choose the targets, they do not judge the findings, the payout is triggered by a transcript anyone can replay rather than by anyone’s verdict, and the purse cannot be withdrawn by the party it is spent on. Whether that last comes by levy or by an agreement made irrevocable is for legislators and laboratories to settle, and this book prefers the former, for the same reason that every publicly listed company funds an audit its management is barred from appointing. A network rewarded for discovery and paid on proof has no stake in what it turns up, which is what makes the miner the right model. How that is engineered is for founders to work out, and for the market to incentivise. Nor must every act of cognition be checked, any more than every taxpayer needs auditing. A sample drawn at random from what the systems did, across providers and unannounced, holds all of the output to the reference value at a fraction of the cost. What the sample cannot do is settle a single case, so it measures the setting and its drift rather than any one outcome, which is the scale at which this section has said the capacity is lost. Three things keep it honest. The population sampled is fixed before the draw, so that the party measured cannot choose its good days. The draw is made by someone else. And the records drawn belong to the people who chose to submit them, which portable memory, one of the tests proposed later, makes possible and which no laboratory could grant on their behalf. Where a check sits is a design question with real content, and not mine to settle. Which checks can be moved into the execution path, at what cost in latency, and which have to sit alongside it and accept being slower, is the sort of question founders and the market are better placed to answer, and I should decline to. What the book owes you is the criterion. What to build against the criterion, and at which parameters, is engineering, and it will be revised as the parts improve, which is the correct relationship between a constitution and an implementation. *** Nick Land’s book on Bitcoin, Crypto-Current, makes a claim close to mine. He argued that Bitcoin has turned rigorously constitutional governance, a problem the old liberal order could never solve, into a technical capability under automatic propulsion, and that the industrialisation of government follows, with regimes becoming, potentially, business outputs. Land reached this chapter’s thesis before I did. Constitutional government, understood as a rule that binds without an institution standing over it, stopped being a problem of political philosophy and became a thing that can be built. He is right. The disagreement between us is now much sharper than it would otherwise be, because it is no longer about whether the mechanism exists. It is about two things only: what reference value the mechanism should hold, and whether voice survives. On the reference value, nothing in the mechanism selects one. The point made at the start is this: a governor enforces and does not decide. Land’s constitution is proprietary and disciplined by exit: a regime is a product, its customers may leave if they dislike it, and that leaving is all the say they have. Mine is a commons, and the reference value it holds is that no party may wield uncheckable power over the rest, which is the non-domination this book has been building towards. Both are reference values. The mechanism will hold either with perfect indifference. So the choice between them is not a technical question at all, and no amount of engineering will settle it — which is a point in favour of my argument and against the tendency, common on both sides of this quarrel, to think the technology has a politics of its own. On voice: a regime that is a business output has customers, and they may leave but may not speak. That is not an incidental feature of Land’s position. He states it as a virtue, and it is the point at which his argument and mine become incompatible rather than merely differently emphasised. Land also posed the obvious question about proof-of-work — could all that computation not be doing something useful? — and answered it in the negative, on principle, holding that the perfect pointlessness of the generation procedure is a feature rather than a bug. The reasoning is that if the work had value elsewhere the cost signal would be corrupted, since the price of an attack would be discounted by whatever the attacker could earn from the computation, and the security would degrade accordingly. But Land is objecting to one way of meeting the need for checkers who cannot be cheaply impersonated and whose interests diverge. Proof-of-work makes identity expensive by making work meaningless. Any other mechanism that makes identity expensive and keeps interests divergent satisfies the same condition equally well, and would do so whether or not the expense took the form of meaningless computation. So his argument establishes something about proof-of-work but nothing about governors generally. That is the artefact-versus-recipe confusion this chapter exists to correct. He has mistaken one implementation of one condition for the constitutional principle, which is the same mistake made, from the opposite direction, by everyone who thinks the lesson of Bitcoin is that you should own your keys. Land understood the thermodynamics of Bitcoin perfectly well. What he did not do was centre the energy dynamic, since in his account thermodynamics enters largely as an analogy for irreversibility rather than as its substance. The narrow claim is that compute as a contested budget — the question of who decides what the world’s computation is spent on, treated as a political matter with a distribution and a set of claimants — is not in his account. Where compute appears there, it is getting cheaper, which is a fact about supply rather than a question about allocation. That is the gap this book works within. *** There is a gradient running through the three cases, and it is in how much of the reference value the governor itself supplies. In money it supplies all of it: the target is a number, twenty-one million, and the governor makes that figure true. In contracting it supplies the rule but not the result: the target is that whatever the code specifies will happen decided by whoever wrote it. In cognition it supplies neither, because a governor that fixed what a model is for or what a person should think would be teleological subsumption with a constitution. What it holds a system to is not a goal but three facts about purposes: that it has declared one, on record, that the declaration can be checked against what it does, and that the people it influences can change it should they wish. The governor guarantees the question, never the answer. It weakens along the series, which follows from the trade-off named in the contracting section, generality bought at the price of decisiveness, and by the third case, cognition, what is guaranteed is thin indeed: not that you will think well, not even that the systems will serve you, only that the means of contesting what they are for remain in existence. The reply to the objection that this is too thin to be worth having is a claim about what constitutions are for. No constitution has ever guaranteed outcomes. What a constitution does is keep outcomes contestable. Free speech does not ensure anybody speaks well, but it makes silencing answerable. The vote does not guarantee a wise choice, but it renders power removable. Due process does not produce correct verdicts, but it makes wrong ones challengeable by the person they fall on. Every constitutional guarantee has this thin, procedural shape, and every one of them has turned out to be worth more than the substantive guarantees offered instead. So what is being proposed for cognition is a constitutional guarantee of the ordinary kind, in a domain that at present has none at all. It has to be constitutional rather than left to the market, which is where David Sacks would leave it, for the reason the capacity condition gave: trade disciplines bad actors through the judgement of the people doing business, but it cannot protect the faculty it depends on. The requirement therefore belongs in the rules every polity shares and not in someone’s terms of service. This is also the point at which the transfer argument and the capacity condition meet. Both end in the same result: the constitution creates the conditions and cannot compel their use. A polity that has the means of contesting a thing and declines to use them has not been failed by its constitution. It has made a choice, which is all a constitution can preserve for anyone. *** There is one place this argument will be fought, and it is not a philosophical objection. The condition that the checkers cannot be cheaply impersonated may be challenging in cognition specifically, and the reason is not money. A Bitcoin full node needs nobody’s permission: the ledger is public and the rules are in the open. Checking a frontier model depends on access to its maker controls: the weights are private, serious evaluation needs entry the laboratory can grant or withdraw at will, and most people able to do the work are former employees of the firms under scrutiny, sharing their trade, their outlook and their future employers. A purse does not cure that. Only a rule about access does, and the tests proposed later supply one. So the bet is empirical and it is about structure rather than price: whether the capability to check can be held by parties with different interests, or it stays inside the building whatever it costs. Should the tests be adopted, outside checkers would gain entry by right, not by favour, and the wager could then be lost on its own terms. If they keep missing what the makers’ own teams catch, or verifying a model becomes as dear as building one, the condition of divergent checkers fails in cognition and the method does not carry over. So: Bitcoin holds a single reference value. It re-imposed scarcity on a substrate that had abolished it, and it proved that a rule can be written in code and held at the speed of what it constrains — the machine — with no institution above it and no hand on the tiller. It is not a model for a society, nor is it a currency I am urging anybody to adopt. What it demonstrates is a method that transfers under four stated conditions which I have now checked in three domains and found to hold completely in one, partially in the second, and unevenly and with one serious doubt in the third. Scarcity is not the only thing subsumption dissolves and money is not the only domain that needs a governor. Provenance, personhood, cognition and the conditions of plural authorship all meet the same solvent and will want a governor cut to their own shape, and each will have to be taken through the same four tests. It is proposed anyway because the doubt is empirical rather than principled, because the domain at present has no governor of any kind, and because a method that names the condition on which it fails is worth more than the alternatives on offer, which identify none. None of this is lost on the priesthood, and it is something my venture capital firm and I share with them. The same venture treasuries that pour capital into the frontier labs also often fund these same primitives — the distributed ledgers, the tokens, the cryptographic rails — and they are right to, because the tools are real and powerful. What differs is what the primitives are bound to. In their assembly the cryptography is plumbing for the machine economy, a way to let autonomous systems transact and settle without human friction: incentives with no reference value above them, the pure accelerant this book described earlier. A heap of ledgers and tokens is no more a governor than a pile of flyballs and linkages, until it is bound to a reference value, just as Watt’s loose parts became one only once they were arranged to hold a speed. The priesthood reached for these and wired them into the deletion. The same parts, wired instead to a reference value a community has deliberately chosen, become a governor. Same primitives, opposite constitution. The mechanism holds whatever it is given and does not care what that is. So the next thing the argument owes is the purpose the reference value serves: what, in the end, all these governors are for, and by what right anybody sets it. That is not an engineering question and no protocol will settle it. It is the oldest political question there is, arriving in a new place.