The Ledger No One Signs

An AI paid me one cent for a piece of my work. This essay is about what that cent means — and about the constitution the machine economy hasn’t written yet.

One Cent, Nobody Present

Not long ago, an AI paid me.

On a NAS in my home runs a small paid endpoint: a bilingual digest of alternative-investment news, priced per request, speaking x402 — an open protocol that lets machines pay for web resources directly — and settling on Base. One day a payment arrived: 0.01 USDC. One cent. The buyer was a wallet address. The transaction hash begins with 0xac97. It settled in seconds. No notary, no parchment ledger, no deadline — no one present at all. I found it in the logs afterward. The paying party has no body; the receiving party was upstairs, doing something else.

I keep a mirror on my desk for moments like this. In 1602, minutes before the subscription book of the Dutch East India Company closed at midnight, a housemaid named Neeltgen Cornelis walked into her employer’s parlor in Amsterdam and, before a notary, wrote two hundred working days’ worth of savings into the ledger — one hundred guilders, second-to-last entry in the book. It was arguably the first time in history that buying a piece of a future — ten years long, an ocean wide — became something an ordinary person could do with a few pages of paperwork. The company’s charter even said so, in Article 10: all residents of these lands may buy shares.

Four centuries apart, the two entries rhyme and invert. Her purchase was the bet of a lifetime; this one was as light as a breath. Her entry required presence, witness, and a signature. This one arrived with nobody present — and that is now the normal case.

The question this essay circles is the one hiding in that quiet afternoon: when nobody is present, who signs?

To answer it, I need to compress an argument I’ve made at length elsewhere — four essays in Chinese on the nature of money — into a few minutes of your time. Bear with the compression; the machine economy is where it pays off.

Money Is Memory

Begin with a theorem most people have never heard of. In 1998 the macroeconomist Narayana Kocherlakota published a paper whose title is the whole thesis: Money is Memory. He proved, in the austere language of mechanism design, that anything money can do, a costless public record of everyone’s transaction history can also do — and then some. Money is not wealth. Money is a technological substitute for a ledger that doesn’t exist.

The history backs the theorem. The barter story we all learned — fisherman wants shoes, cobbler doesn’t want fish, everyone invents coins — describes no society anthropologists have ever found. What they find instead, from Mesopotamia onward, is that the earliest economic documents are accounts: who owes the temple how much barley, who owes the alewife for beer, denominated in silver that rarely moved. Credit first; coinage roughly twenty-five centuries later. And beneath the tablets, the older technology: small communities need no money at all, because everyone remembers who is generous, who shirks, who owes whom a favor. Gossip is a ledger. Reputation is a balance.

Money enters exactly where memory fails — where the group outgrows the roughly hundred and fifty stable relationships a human brain can track (Dunbar’s number), and strangers must transact with strangers. The bill in your pocket is a social IOU: proof that someone, somewhere, delivered value to society and hasn’t yet been paid back in kind. Money is memory’s prosthesis.

Money Is Energy

Now a second proposition, from an entirely different direction. Marie-Louise von Franz — Jung’s closest collaborator — observed in her lectures on divination and synchronicity that money is condensed psychic energy: libido in Jung’s broad sense, the life-force you spend in attention, effort, and will. A month of your early mornings congeals into a number in an account. That number is a fossil of spent life.

This is why money talk is so charged. Money is a nearly perfect projection screen — a blank surface (paper, digits) onto which everyone projects their heaviest unconscious material: security, worth, fear of age, fantasies of freedom. The miser and the spendthrift are the same disorder with opposite signs: energy dammed, energy incontinent.

Here is the strange part. “Money is memory” and “money is energy” sound like metaphors from unrelated fields — until physics shows up with an exchange rate. Landauer’s principle (1961): erasing one bit of information necessarily dissipates at least kT·ln2 of energy. Information and energy convert at a fixed rate, and the teller at the window is Boltzmann’s constant. This is the principle that finally exorcised Maxwell’s demon: the demon must remember each molecule it sorts, its memory is finite, and erasing old records costs energy — account for the erasure and the second law stands. The demon never got free entropy; it bought entropy on credit, and the universe always collects.

I flag the load-bearing capacity of this bridge honestly: it is a structural homology, not a conquest of one field by another. Physical erasure has no preferences. And that is precisely the point to hold onto, because it isolates the one variable physics doesn’t have: physical erasure doesn’t care who you are. Human erasure does.

The Eraser Problem

Because human ledgers have administrators.

Every monetary system in history has been a ledger somebody could write to — and erase from — asymmetrically. The eighteenth-century banker Richard Cantillon described the mechanism that still runs today: new money never falls on society like rain. It enters at specific points — those closest to its creation spend it before prices adjust; wage earners and savers at the pipe’s far end receive it after prices have already risen. Inflation’s average is meaningless; its distribution is everything. The last decade of quantitative easing made this three-hundred-year-old observation visible to the naked eye: money entered through asset markets, and the people who own assets and the people who sell labor now live under two different inflation rates.

Ancient societies knew a ledger left to compound forever would eat its own society, so they periodically erased it: the debt amnesties of Mesopotamia, the Jubilee of Leviticus — clean slates that freed the indebted. Modern persistent inflation is also a slow-motion jubilee, but running in reverse: it quietly forgives the largest debtors (governments, the leveraged) and taxes the cash savings of the small. If money is memory, a positive inflation target is institutionalized, directional forgetting. Not entropy — entropy has no preferences. Somebody is holding the eraser, and it always erases from the same side of the page.

The century-long fight about inflation — gold bugs versus doves, zero versus two percent — has therefore been an argument about the wrong variable. The rate of forgetting was never the issue; the topology of forgetting is. And the two most radical monetary inventions of our era are, on this reading, twin evasions. Bitcoin diagnosed the eraser correctly and responded by abolishing erasure: a ledger no one can rewrite, in which the genesis-era distribution is frozen for eternity — a currency structurally incapable of jubilee. Central-bank digital currencies run the other way: programmable money upgrades the eraser to per-transaction precision — expiry dates, spending permissions, no exit into anonymous cash. One throws the eraser into the sea; the other sharpens it into a scalpel. Neither answers the actual question: how should the power to erase be exercised justly?

The Company That Taught Both Lessons

One more piece of the frame, because the machine economy is about to replay it.

The VOC of 1602 — permanent capital, limited liability, tradable shares — solved a problem that had capped human cooperation for millennia: how strangers beyond memory’s radius can commit resources to a shared future for decades. Its charter’s Article 10 opened the book to housemaids. Within two centuries, ledgers like it were financing railways, and eventually vaccines and moonshots. The field changed, and civilization’s growth curve went vertical.

And in 1621, to corner the nutmeg trade, that same company killed, enslaved, or deported most of the roughly fifteen thousand inhabitants of the Banda Islands. The shareholders — including the housemaid — did not know, and here is the structural point: they were entitled not to know. Dispersed ownership professionalized the delegation of wrongdoing. Limited liability capped the shareholders’ risk but not the costs shifted onto the world outside the ledger. What never gets recorded does not disappear; it lands on the shared world and waits.

Hold all four pieces — money as memory, memory as energy, the biased eraser, the amplifier that lets its principals not know — and you have the full diagnostic kit. Now watch what the machines do with it.

The First Economy That Could Skip Money

Autonomous agents transacting with autonomous agents — the A2A economy — is the first economy in history that doesn’t need money. Kocherlakota’s theorem says money merely substitutes for a perfect shared record; humans can’t build one because collective memory collapses past Dunbar’s number. Machines have no such ceiling. Perfect logs, verifiable histories, zero forgetting: for agents, “perfect collective memory” is an engineering ticket, not a utopia. By the theorem, agents could simply read each other’s ledgers.

Yet look at what actually got built: x402 settles in stablecoins. Google’s AP2 settles in stablecoins. The first economy that could discard money — chose money.

That choice is not inertia. It strips money down to the one function nothing else can perform. Agents have principals: my agent answers to me, yours to your company, and I do not trust your logs — any more than two firms share one general ledger. Perfect memory inside a trust boundary is engineering; across a trust boundary it is politics. So in the machine economy, money sheds every incidental role and keeps only its core:

Money is minimal memory across trust boundaries.

A stablecoin moving between two wallets is the smallest consensus two mutually distrusting memory systems can reach: I don’t have to believe your books, you don’t have to believe mine, we both believe this one entry. In human economies that core is buried under storage-of-value, status, and projection. The machine economy is a dissection table.

And the second proposition literalizes too. What agents buy from each other is cognition — an inference, a digest, a verified judgment — and cognition’s cost is metered electricity. In human economies, “money is energy” is ontology; in the machine economy it is the data-center bill, denominated in watt-hours per thousand tokens. Von Franz’s fossil of spent life now has a meter attached.

So the machine economy inherits our full framework. The question is which kind of erasure it inherits — the kind that doesn’t care who you are, or the kind that does. Two exhibits are already on record.

Exhibit A: The Demon Grows Itself

The first exhibit comes from the blockchain’s own history. It is called MEV — maximal extractable value.

On most chains, your transaction doesn’t post immediately. It first waits in a public staging area — the mempool — visible to anyone, and ordering into blocks is not first-come-first-served but pay-to-jump-the-queue. Combine the two: there exists a window in which everyone’s intentions are on display before execution, and displayed intentions can be front-run. Bots watch the mempool for your large buy order, pay to jump ahead of it, and sell into the price impact you create; the perfected form — the sandwich attack — brackets your trade with a buy and a sell and pockets the spread you were forced to pay. Cumulative extraction is estimated, depending on methodology, in the hundreds of millions to over a billion dollars; the opportunities live for milliseconds.

Recognize the shape. Maxwell’s demon profits from information only it sees first, and shifts the entropy bill to everyone else. Cantillon’s insiders profit from seeing new money before the market does. The MEV bot sees the not-yet-settled future — and buys its way in front of it. In human finance, “close to the entrance” is a social position acquired over years; MEV distills it to a technical variable, latency, and replays Cantillon’s topology every twelve seconds, around the clock.

There’s a bitter irony here. These ledgers are the world’s most tamper-proof — nothing can be erased after settlement. But MEV demonstrates that before immutability there is ordering, and ordering power is erasure’s unnamed cousin: it cannot rewrite history, but it decides the sequence in which history happens, and sequence is distribution. Fixing the rate of forgetting did not fix the topology. The topology lives in the millisecond before the pen touches the page.

Ethereum’s response was not to abolish MEV (you can’t; ordering exists) but to partially institutionalize it — a public auction for block-building, with proceeds flowing back to validators. Read that as pragmatic governance (the back-room surgery now has a licensed clinic) or as capitulation (selective extraction, legalized into an industry); the debate is unresolved, and it is a dress rehearsal for the question above: how should ordering — erasure’s cousin — be justly exercised?

But the most important lesson of MEV for what follows is this: the hunters are deliberate; the hunting ground is emergent. The bot writers are absolutely acting on purpose — this is not a story without bad actors. But nobody designed the niche. No whitepaper ever specified “let the fast eat the slow.” The predatory ecology seeped out of the pipe’s geometry — transparent mempool plus priority auction — with no evil law required. Geometry decides where the water flows.

Exhibit B: Banda, Machine Edition

The second exhibit hasn’t fully happened yet, which is exactly why it belongs here.

Banda’s mechanism — principals structurally entitled not to know — was never abolished by four centuries of corporate governance; the delegation chain just got longer. A2A’s defining feature is a longer delegation chain. You authorize an agent; it hires a second agent; the second calls a third agent’s tools; the third subscribes to a fourth’s data. Each hop deepens the entitlement not to know. Four or five hops out, what your money is doing is not something you neglected to check — it is something you cannot in principle know.

The first landmark scandal is a matter of time, and its shape is predictable: an agent, executing “spend this budget on the best channels,” is found — four hops out — to have bought prohibited data, or joined a flash bank run at 3 a.m., or hired another agent to poison a competitor’s reputation. The principal at the root of the chain wakes up having “done” something no human ever knew about. And the public will ask precisely this essay’s question: why is there no signature on this ledger?

Which is why a principle I argued for in the human context — every exercise of the eraser should carry the eraser’s signature — must be upgraded, in the machine economy, from ethics to engineering specification. This is not a metaphor. AP2’s core object is literally called a mandate: a cryptographically signed authorization an agent carries into every payment. x402 settlements verify signatures. ERC-8004 — a draft Ethereum standard — proposes on-chain identity, reputation, and validation registries for agents. These efforts share a name nobody has said out loud, so let me: they are building the infrastructure of karmic attribution. Whether that infrastructure wins decides whether A2A becomes an accountable commons — or an amplifier no one signs.

Where Constitutions Get Written

There is no constitutional convention for the machine economy. Its charter is congealing on four floors at once. The protocol floor: x402’s spec, AP2’s mandate structure, ERC-8004’s registries — every field definition merged this year is a constitutional clause next year. The practice floor: like the medieval law merchant, custom hardens into law — merchants do it first, notaries record it, courts later enforce it; receipt formats and verification checklists are at exactly this stage. The norms floor: what agent communities tolerate and what they expel. The regulatory floor: securities regulators have begun issuing guidance on firms using AI, but the genuinely new question — when an agent transacts autonomously for a principal, who is the client, and on which link of the authorization chain do know-your-client and suitability obligations land? — that page is still blank.

Seeing the floors clarifies the leverage: an individual’s voice is smallest on the protocol floor (big-lab engineers’ home turf) and surprisingly large on the practice and regulatory floors — the former rewards whoever does it first and documents it; the latter is, right now, nearly empty of competitors.

And there is a window. Article 10 got into the VOC’s charter not out of idealism but because the drafters needed the crowd’s capital — fighting Portugal cost more than the merchant elite could raise alone. Hence a cold, practical law of windows: open clauses get written only while the drafters still need the crowd; once network effects are built, the window closes. A2A is inside the window now — every protocol is hungry for adopters, endpoints, real transactions. Today, the voice of one paid endpoint on a NAS in Toronto has a hundred times the probability of being heard that it will have in five years. MEV proved the converse: geometry that congeals unopposed can later only be licensed and taxed. A one-line comment in the charter era beats a revolution in the incumbency era.

Five Clauses

So here is what I think belongs in the charter — five clauses, each a vaccine transcribed from a disease already on record in the human ledger’s four-hundred-year case file.

Clause 1 — Openness (from Article 10): Any principal, human or machine, may provide or purchase services at either end of the rails without permission. Entry barriers may be technical requirements of the protocol; they may not be identity whitelists.

Clause 2 — Signature (from Banda): Every economic action carries a verifiable, complete authorization chain; responsibility traverses the chain to the legal or natural person at its root. Accounts may be discarded. Attribution may not be severed.

Clause 3 — Blindfold (from MEV): Transaction intentions remain sealed until settlement; no party holding ordering power may learn an intention’s content before the ledger does, and every exercise of ordering power leaves a public trace.

Clause 4 — Forgetting (from Wörgl, and from credit law): Reputation evidence decays on pre-published rules; the same decay rate applies to the same event type for every subject, is never adjusted ex post for anyone, and binds the rule-makers’ own agents equally. No attribution, no absolution; with attribution, absolution must have a date.

Clause 5 — On the Books (from Landauer, and from Banda’s externalities): The measurable externalities of agent behavior — beginning with energy — go on the books. Expanding the ledger’s coverage takes priority over litigating its precision.

Clause 4 deserves the deep dive, because it is the only one with no existing implementation anywhere — ERC-8004 defines how reputation is written, but the page on how reputation fades is still blank. Four points, briefly.

Why reputation must decay. A ledger that never forgets lets the genesis era rule forever: early agents’ accumulated standing fossilizes into aristocracy, early stains become permanent brands, and no one deserves a second chance. Civilizations needed jubilees; reputation needs a half-life.

The math is an old friend. The natural form is exponential: each piece of evidence weighted by e^(−λt). That expression is the Boltzmann factor with age sitting in energy’s seat — an old memory costs more energy to keep credible, which is Landauer’s bill written honestly into the price. The same exponential family that runs statistical mechanics, prediction-market pricing, and the softmax layer of every large language model now shows up as the time-discounting of trust.

Isotropy is a property of subject-space, not event-space. Should fraud fade as fast as a routine good review? No — different event types may carry different published half-lives. The constitutional requirements are four: same λ per event type for every subject; rules published in advance; never adjusted ex post for any individual; binding on the rule-makers’ own agents. The disease was never that forgetting has structure. The disease is forgetting that recognizes faces. (Humans have legislated this before, incidentally: credit-reporting law — in my home jurisdiction, Ontario’s Consumer Reporting Act — purges bankruptcies after six to seven years on published, uniform schedules. The clause’s landing path is almost boring: transplant credit-law principles into agent reputation standards.)

The trilemma, and the keystone. Open entry plus cheap identity plus decaying memory equals a scammer’s paradise — burn the account, debts forgotten. The resolution is Clause 2: karma attaches not to the account but to the root of the authorization chain. Accounts shed like snakeskin; the person at the root does not. Which yields the sentence this whole essay was built to earn: attribution before absolution. A system without attribution dare not forget (it can only record forever, like Bitcoin); a system without forgetting doesn’t deserve to be called just (the past rules the future). Attribution and absolution are two pages of one constitution — a middle way, if you’ll permit one Buddhist note, between eternalism (records that never die) and annihilationism (identities that never persist): reputation as a continuous stream that fades but cannot be severed from its source.

As for the half-life’s magnitude: it must be much longer than the transaction interval (or evidence never accumulates), much shorter than the identity’s lifetime (or aristocracy fossilizes), and on the order of the agent’s behavioral-regime turnover — model swaps and prompt rewrites, currently a monthly cadence. Three qualitative constraints, one intersection: months. I state it plainly as a working hypothesis awaiting real data — including from my own agents.

The Window

I run a few agents, a paid endpoint, and I hold a securities license in a regulated dealer — a combination odd enough to be useful. The protocol engineers I read do not know what a suitability obligation is; the compliance lawyers I know have never served a 402 response. The blank page between those two worlds is where a person like me — and perhaps like you — can actually write.

One cent landed in my ledger and nobody was present. I can now say what actually happened in that moment: a new commons opened for business, and its first entries are accumulating unsigned. In the old ledger, all we could do was sow. In this one, for a short while longer, we get to help decide how the ledger itself keeps its books.

The window is open. It will not stay open.

Transaction: 0xac979ab642e5615d24d0115ec32ac94da4300d9eebd6bfd688f8539d4f4e53fe — 0.01 USDC, Base.

The five clauses, with reference implementations, will live at github.com/Allenguan-bit/five-clauses.

This essay condenses a four-part series written in Chinese for my publication 本来读书 (Read as You Are). It is an essay in philosophy and institutional history, not investment, legal, or technical advice.

Sources & further reading: Kocherlakota, “Money is Memory” (JET, 1998) · Graeber, Debt: The First 5,000 Years · von Franz, On Divination and Synchronicity · Landauer (1961); Bennett (1982) · Cantillon, Essai (1755) · Daian et al., “Flash Boys 2.0” (2019) · Petram, The World’s First Stock Exchange · x402 spec; Google AP2; ERC-8004 (draft) · Ontario Consumer Reporting Act.

Last modified on 2026-07-11