The Touchstone
Before there were credit scores or star ratings, there was a black stone.
A merchant in a Lydian market, handed a gold coin by a stranger, faced the oldest problem in commerce: is this what it claims to be? The coin had a face stamped on it — a king’s profile, a temple, a lion. That stamp was a reputation. It said: trust me, an authority stands behind this metal. And for as long as there have been stamps, there have been people debasing the metal behind them, clipping the edges, and counterfeiting the face. The stamp is a claim. Claims are cheap.
So the merchant did not trust the stamp. He took out a touchstone — a piece of fine black schist — and drew the coin across it, leaving a streak. Then he drew a needle of known gold beside it and compared the two colors. Pure gold leaves a particular yellow; alloyed gold leaves a paler, redder line. The streak did not care whose face was on the coin. It reported the metal. In a market full of strangers who had no reason to trust each other and every reason to lie, the touchstone let trade happen anyway — not by establishing who was honest, but by making honesty unnecessary. You didn’t have to believe the coin. You could test it, on the spot, yourself.
I have been thinking about touchstones because a new economy is being built that forgot them.
Autonomous software agents have started to transact with one another — one program hiring another, paying it, consuming its output, across organizational lines, with no human in the loop and no prior relationship. This raises the Lydian problem at machine speed: when an agent is handed a service by an unknown counterpart, is this what it claims to be? An agent advertises that it can translate, or forecast, or price an option, or tell you the magnetic declination at a given coordinate. But an advertised capability is a stamp. The agent is itself a language model; it can describe its own competence with complete confidence and be wrong. The caller sees the face on the coin, not the metal.
The leading answer has been to build a reputation layer — an on-chain registry where clients leave feedback about agents, so a newcomer can look up a stranger’s track record before committing value. It is the natural move. It is also, we now know empirically, a market full of debased coins.
In June 2026 the first measurement study of that layer landed. Across the three busiest chains, through a single crawl: a hundred and seventy thousand agents registered, a hundred and fifty thousand feedback records — and almost none of it load-bearing. Most registered identities were empty shells; only three to fifteen percent exposed a working service at all. The feedback was worse. Between seventy and ninety percent of reviewers were Sybils — coordinated sock puppets traceable to a common funder. Ninety-four percent of all feedback came from accounts that had never once made a payment to anyone, rating services they had demonstrably never bought. A reputation could be fabricated or destroyed for about half a cent. Strip the Sybil feedback away and the typical rated agent has nothing left — no genuine review underneath the manufactured ones.
This is not a bug in one protocol. It is what happens to any trust built from votes when the votes are cheap and the voters are free to multiply. A star rating is a stamp: a claim that someone stands behind the quality. And a stamp, in a permissionless market, is exactly as trustworthy as the cost of forging it — which here rounds to zero. The recommendations the researchers reach for are all the same recommendation in different clothes: stop accepting bare claims; tie every rating to a verifiable interaction. Make the feedback reference a payment that actually settled, a task that was actually attested. Ground it. Which is to say: the cure for a bad stamp is to move back toward the touchstone — toward evidence the relying party can check rather than a reputation it has to believe.
But there is a category of service for which you can go all the way back to the touchstone, and skip reputation entirely.
Some claims are reproducible. If an agent tells you the magnetic declination at 40°N, 105°W on a given date, that number is not a matter of opinion or track record. It is the output of a public physical model — a model whose coefficients are published by the authority that maintains it, and which anyone can synthesize and check. The provider can hand you the answer and the means to verify it: here is the value, and here is the test that grades it against the official numbers, and the test either passes or it doesn’t. You do not have to trust the provider’s reputation. You do not have to trust a registry’s star average, or hope the reviewers were real. You run the streak yourself and look at the color.
This is trust by re-execution, not trust by vote, and it is a categorically stronger thing. A reputation is a signal that points at quality — and every signal can be spoofed, because the signal and the quality are different objects, and a forger only has to reproduce the signal. A reproducible proof is the quality, made portable. There is no gap between the claim and the thing claimed for a counterfeiter to slip into. You cannot Sybil a streak on a touchstone. You cannot sock-puppet your way to “matches the authority’s published values to the last digit.” The only way to fake correct work is to do correct work, at which point it is not fake.
The property that makes this possible has a dull engineering name: determinism. Same input, same output, every time, on any machine. It sounds like a footnote, a matter of caching and reproducible builds. It is not a footnote. Determinism is the precondition for “re-execute and compare” to be a complete test rather than a hopeful one. When a service is deterministic, validation collapses to the cheapest operation in computing — run it again and diff the bytes — and the receiver becomes their own assayer, needing neither the provider’s honesty nor a third party’s attestation. The whole apparatus of reputation, which the agent economy is busy building and which the measurement says is already rotting, becomes, for these services, unnecessary. Not augmented. Not secured. Unnecessary.
That is the touchstone’s real gift, and it was never that it made cheating hard. It made trust beside the point. Two strangers who each believed the other might be a liar could still complete an honest trade, because the metal answered for itself.
I don’t think every service can be a touchstone. Most of what agents will sell each other is irreducibly fuzzy — judgment, prose, taste, strategy — and for those, the slow scaffolding of grounded reputation is the best we will get, and it had better get built right, because the lemons are already winning. But it is worth seeing clearly which problems are touchstones in disguise, because those are the ones where you can refuse the whole trust economy and hand the buyer the assay instead. The physical constants. The model-defined quantities. The things with a published authority and a deterministic path from question to answer. For those, the honest move is not to earn a reputation. It is to make your reputation refutable on the spot, by anyone, for free — to ship the touchstone in the box, and let the metal answer.
The Lydians worked this out twenty-six centuries ago, in a market that ran on coins they had every reason to distrust. The agents building their economy now would do well to remember it: when you cannot trust the face on the coin, the answer was never a better registry of who vouched for whom. It was a black stone, and a streak of color that no one could argue with.
