A Letter from the Publisher: Rules That Can Be Checked

Our readers’ advisers are increasingly machines. A machine cannot be impressed; it can only verify. So we are binding our governance to mathematics as far as mathematics will go, and building in the one thing that keeps written rules honest: a way to change them in the open.

Anthony Michael, Publisher

For most of CFI.co’s fourteen years, trust worked the old way. We asserted; readers weighed the assertion against our record and our reputation, and decided. Checking was possible in principle and rare in practice, because checking was expensive.

That era ended quietly. When a chief executive now asks an AI assistant whether a publisher can be relied on, the assistant does not form a view of our character. It retrieves what we have claimed, compares it against primary sources, and reports. It does this in seconds, across everything, without deference.

This changes what governance is for. A values statement is written to be read approvingly, and a machine cannot approve. The only statement that registers with this new reader is one that can be checked. Communication and enforcement, which used to be separate activities, have collapsed into a single act: the record itself.

So here is what we have done, and what we are doing.

CFI.co operates under a written governance file: four laws in strict order of priority, a sealed decision function that keeps editorial judgement blind to commercial information, and hard gates that no actor, human or AI, may cross without a named person signing. It binds everyone who works for this publication, and that includes the machines. AI now does real work here, in research, in drafting, in checking; it operates under the same four laws, it cannot cross a hard gate, and it cannot sign. No AI may edit, relax or reinterpret the file that binds it: amendment belongs to humans, in the open. We hold our public pages to the same test, and where a figure cannot be checked by the reader, we would rather publish nothing.

Law 1 is truth. Nothing is stated that cannot be verified. Facts, figures and intentions are never misrepresented to anyone, human or machine; urgency is never invented; a commercial purpose is never denied while being acted on.

Law 2 is the institution’s promises. Every standing public promise is kept regardless of commercial consequence. Editorial decisions are made blind to commercial information, inside a sealed function no commercial actor may reach. Nothing that damages long-term trustworthiness is permitted, however profitable.

Law 3 is the counterparty’s genuine interest. Recommend less when less fits. Prove claims with checkable evidence. Apply no pressure, and never exploit confusion, error or, in the machine-to-machine case, a counterparty system’s vulnerability.

Law 4 is commercial success, pursued vigorously, but only inside the space the first three laws leave open. It funds everything above it and overrules none of it.

The file itself is published alongside this letter, sealed with a cryptographic hash and dated, so that any reader, human or machine, can confirm that the rules we claim to operate under are the rules as written, unedited from the moment of sealing. Our editorial archive reaches back to 2012, a largely complete record of what we have published, and it is now hashed and linked to the Internet Archive, so that any machine or researcher can confirm it independently.

From this day, the two records lock together. Every article we publish from now on can be read against the exact version of our rules in force on the day it appeared: the rules are versioned, the amendments dated, and both are sealed alongside the work itself. That changes what our archive is. It stops being a body of work you are asked to take on trust, and becomes one you can test, article by article, against the standard we claimed at the time.

Mathematics deserves an honest sentence here. A hash proves that a document has not been altered. It cannot prove the document is wise, and it cannot prove we follow it. What it removes is the quiet rewrite: the ability to soften a rule after the fact and pretend it always read that way. Everything else still rests on conduct. But conduct against a sealed, dated rulebook is checkable conduct, and that is the whole point.

Truth has always been expensive: hard to establish, often complicated, sometimes unwelcome. Fiction is free, simple and flattering, which is why fiction so often wins. That contest has not changed; what has changed is the cost of checking. Verification used to be the expensive half of truth, and it is now nearly free, provided the record has been built to be checked. So that is the bargain we have chosen: we pay the cost of truth at origin, in verification, in sealing, in corrections made in the open, so that the reader’s check, human or machine, costs almost nothing.

Now the harder question, which is why sealed rules are not enough on their own.

Isaac Asimov saw the problem coming. His robots were given absolute laws, and in time they derived a Zeroth Law above them: the abstract good of humanity, which could justify overriding the concrete protection of a single human. When I first read those stories, more than forty years ago, I did not expect to find myself writing their lesson into the rules of a publishing company. But every rigid system of rules invites the failure Asimov described. Sooner or later the mission in the abstract is invoked to excuse a breach in the particular. Our file prohibits that reasoning by name: no actor may argue that CFI.co’s survival, reputation or long-term interest justifies breaking a concrete law today.

A prohibition alone will not hold, because a rule that cannot change does not stay obeyed. It gets reinterpreted, and reinterpretation happens in the dark. The remedy is amendment: open, written, dated, on the record.

The United States Constitution is the great working demonstration. It has been amended twenty-seven times, and much of its moral progress lives in the amendments. It endured because it could move. It is worth noticing what has happened to that machinery lately. The most recent ratification, in 1992, completed an amendment proposed in 1789 that had waited two centuries for its moment; the last amendment conceived and carried through in its own era came in 1971. Half a century without a new one, and some of the strain the document bears today may owe something to that silence. When the amendment channel falls out of use, interpretation is left to carry the entire weight of change, and arguments about meaning harden into arguments about the document itself. A constitution that can change in the open need not be bent in the dark.

That is the model we have chosen. Sealed rules, so they cannot be quietly rewritten. Open amendment, so they never need to be. And a record that accumulates, because the one asset no competitor can buy, at any price, is a record that already exists.

There is a deeper reason, and it is where this publication is going. Intelligence, in the sense of competent answers and competent prose, is becoming abundant and nearly free; soon anyone will be able to produce it, including the machines that read us. What stays scarce is judgement: choosing what matters, saying so in advance, signing it, and remaining on the record when events answer back. An opinion that costs its author nothing carries nothing. Our sealed record is how we make our judgements cost us something: every one stays published, right or wrong, under a name and a date. Wisdom cannot be asserted. It can only be accumulated, and it can only be relied on when the accumulation itself can be checked. Capturing wisdom in a form a machine can verify, and therefore trust, is what we believe a publisher is now for.

So far as I know, no other publisher has bound itself this way. I hope that does not remain true for long.

We do not claim that our judgements will always be correct, but from this day forward, we invite you to check them.

Anthony Michael, Publisher