The census and the count

Counting people is ancient, and for most of that history it was an instrument of extraction. The Roman census enrolled citizens for taxation and military service; the great medieval land surveys inventoried who held what so that a sovereign could take his due. A ruler counted his subjects the way a merchant counted his stock — to know what he could draw upon. The count flowed one way, from the powerful toward the counted, and the counted had no standing to examine it. Whatever the tally said, it said on the authority of the man who ordered it.

The sharp break comes when a count is made the basis not of what a ruler may take but of how much say the counted may have. The United States Constitution, framed in 1787, requires that an "actual Enumeration" of the population be conducted every ten years, and ties that decennial count directly to the apportionment of seats in the House of Representatives among the states. The consequence is precise and radical: political power is distributed according to a number, and the number is produced by counting people. A state's weight in the national legislature is not asserted, negotiated, or inherited. It is counted, on a schedule fixed in the founding document itself.

Once a count decides something that consequential and that contested, it can no longer be a private act of the counter. It has to become a public instrument that the counted can trust. And so a discipline grew up around the modern census that the old tax-rolls never needed: a fixed reference date, so that everyone is counted as of the same moment and the question "counted when?" has one answer; a defined and published methodology, so that the procedure is a matter of record rather than of the enumerator's discretion; and results published in full, so that the number a state's power rests on is one anyone can see, cite, and dispute. These are not bureaucratic ornaments. They are what convert a tally into a warrant for the exercise of power.

A count that decides who holds power

Consider what the fixed date actually does. Populations move; people are born, die, and travel. A count with no agreed reference moment is not wrong so much as meaningless — two counters can both be right and disagree, because they counted different instants. Fixing the day is the arithmetic equivalent of freezing the rules of a decision at the moment it is made: it makes the count a statement about a determinate state of the world rather than a smear across time. Everything downstream — the comparison between states, the reapportionment, the challenge — depends on there being one thing that was counted.

The disclosed method does something adjacent and just as important. When a number is going to move seats of power between states, the procedure that produced it cannot be a black box, because a black box cannot be audited and therefore cannot be trusted by the party it disadvantages. If the method is on the record — who is counted, where they are assigned, how the ambiguous cases are resolved — then a state that believes it was shortchanged has something to point at. It can argue that the method was misapplied, or that the method itself was flawed. Publication completes the circuit: a result kept secret could be neither checked nor contested, and a count that cannot be contested cannot command the assent of those it governs. The genius of the constitutional census is that it makes the instrument of power also the instrument of its own accountability.

A count that decides how power is shared, made in secret by an undisclosed method, is not a measurement. It is a decree wearing the costume of arithmetic.

Who and how you count is contestable

None of this makes the count innocent. The hardest lesson the census teaches is that the definitional choices — who counts, and how you count them — are themselves consequential and contestable, and no amount of procedural rigor makes them neutral. The American founding document's own original formula for apportionment counted enslaved people as three-fifths of a person, a definitional choice that was a moral catastrophe dressed as a counting rule. Undercounts are not random noise; they fall, persistently and predictably, on particular populations — the transient, the poor, the distrustful — and because the count apportions power and resources, an undercount is a quiet transfer of both away from exactly the people already least able to object.

This is precisely why the discipline matters rather than undermining the case for it. If the definitional choices were buried, an undercount would be invisible — the number would arrive as fact, and the harm would be untraceable. Because the census publishes its method, the choices about who is included and how the hard cases are handled become themselves objects of argument, litigation, and reform. The confidentiality protections that later attached to the individual returns belong to the same logic: people will only submit to being counted if they trust that the record cannot be turned against them, so the legitimacy of the count depends on a promise about how its inputs may and may not be used. A count made in secret cannot command assent; but a count whose definitions are hidden while its total is broadcast is worse, because it borrows the authority of transparency while withholding the thing that transparency is for.

A census discipline for the machine's counts

Now turn the lens. Machine systems produce counts constantly, and route enormous consequence through them. A model tallies the size of a risk population, the number eligible for a benefit or flagged for review, the members of a scored cohort. These aggregates decide how attention, money, and suspicion are apportioned — which is to say they do the same work the census does, allocating consequence according to a number. But they almost never carry the census discipline. The reference moment is unstated; the method is undisclosed; the result is unpublished and, to the people it sorts, unchallengeable. And the definitional choices — what exactly makes someone a member of "the risk population," which edge cases were swept in and which excluded — are buried in a pipeline nobody outside can see.

The failure is not that these systems count. Counting is unavoidable and often necessary. The failure is that they deliver the count the way a medieval sovereign delivered his tax-roll: as an authoritative total from nowhere, produced one-way, offered to the counted as fact rather than as an instrument they may examine. The 1787 answer is that when a number apportions consequence, the counted acquire standing to see how the number was made. That is not a courtesy; it is the condition under which the number can be believed at all.

So the census sets the bar for the machine's counts, and the bar is specific. A consequential aggregate should travel with its reference moment, so we know as of when it is true. It should travel with its method, disclosed well enough that a skeptic can follow how a raw population became a total. Above all it should carry its definitional choices in the open — the inclusion rules, the boundary cases, the assumptions about who belongs to the counted set — because that is where the moral weight actually sits, and where an undercount hides. A count built this way can be scrutinized, and the person it disadvantages can contest it on the record rather than resent it in the dark. This is what a Decision Receipt is for: to make a consequential figure carry the provenance of its own production, so that the people it sorts have the standing the Constitution gave the counted. A number that apportions consequence is a form of power, and power that will not show its method has not earned the assent of the counted — it has only counted on their not asking.

— Dispatches · Summit Cognitive

---