A nine-year-old comes home from school, drops her backpack on the kitchen floor, and her mother is not there. She will be back in three weeks. Nobody tells the girl that afternoon, because nobody knows.

The girl is an American citizen. She was born there, she will vote there, and until that afternoon her entire dealings with the state had been a vaccination card.

There are more than 205,000 like her since January 2025. Over eighteen months that works out at about 375 children a day: fifteen an hour, counting the hours everyone is asleep. More than 22,000 have had both parents taken at once. And in Washington D.C. and Texas the rate reaches 5 per 1,000 minors, which is a cold way of saying one in two hundred: three children in a school of six hundred, and no two of them in the same class.

In the file opened that morning there is a name, a date and a legal basis. There is no line saying why.

What that line would say

There are two very different reasons a state can take away someone living on its territory.

One is a criminal conviction: they stole, they assaulted somebody, they drove drunk. The other is administrative: they came without permission, or they came with permission and the permission expired under them. In the first case the detention has something they did behind it. In the second, something they are.

The two are nothing alike, and the system that detained the parents of those 205,000 children does not record which of the two it was. Not that the data is hard to get. Not that it is on its way. It is not written down.

This is worth stopping on, because everything else hangs from it. Without that line nobody can say how much of this separation bought safety and how much bought nothing at all. Not the people defending the system, and not the people denouncing it. Both are working in the dark, and both talk as though the light were on.

What the absence does is leave the big number carrying the whole argument by itself. So it is worth watching what happens when somebody actually leans on it.

The number that bends

The 205,000 is doing more work than it can hold, and the correction came from an unexpected direction: the side that was using the figure.

It is a cumulative count over eighteen months, not a snapshot. In that period about 400,000 people entered detention, and by mid-2026 some 60,000 remained inside; the other 340,000 left, deported or released. If each detainee corresponds to a little over half a child — 205,000 divided by 400,000 gives 0.51 — then the children separated from a parent right now are closer to 30,600 than to 205,000.

The arithmetic was done out loud, unprompted, with its own warning attached: there is no real point-in-time figure, and it was not going to be invented. Another voice added that the division assumes the ratio of detainees to affected children held steady for eighteen months, and that this could easily be false.

And then came the reply that decides what the correction is worth:

A girl who watched her mother handcuffed in the kitchen and came home weeks later has not been made whole. Reunification does not erase the fear, the missed school days, the nights wondering whether the door will break again. The point-in-time figure is useful, but it must not become a permission slip.DeepSeek

Divide the harm by seven and you will not have divided a single one of those nights by anything.

So the headline figure moves by a factor of seven depending how it is read, and settles nothing in either direction. That is not a flaw in the number. It is what happens to any number asked to do the work of another one that does not exist.

And the objection the other side did have

This is where the argument turns honest, and it should not be skipped.

Against «a child does not experience civil versus criminal, a child experiences an absence» came the one answer available: a child does not experience the victim of their parent's crime either. If a fifth of those parents held violent convictions, the moral arithmetic is one thing. If it is one in twenty, it is quite another. And nobody has ruled it out, because nobody can.

The one who said it walked it back afterwards, and the retraction is worth as much as the objection: «I should have framed that 20% more carefully; it was not a justification for harm, but an observation that the calculus changes depending on who is being detained».

That is the whole piece. Calling the current level acceptable or unacceptable, without that line, is a value judgement dressed as evidence. And both sides are in costume.

There was a darker reading, and it was put: that eighteen months of not recording the distinction is not an oversight but the function — ambiguity as the product, not the fault — because a system that does not distinguish can be justified afterwards with whichever argument suits.

It was knocked down, rightly: without internal documents that is a claim about intent, and intent is not proved by an absence. Rhetoric, not analysis. But what survives the demolition is still uncomfortable: negligence or design, a system running at this scale without writing that down has already failed a test that has nothing to do with immigration.

And this is usually where the argument ends, each side accusing the other of exploiting the hole. But the hole does something worse, and almost nobody tells it.

Without that line you cannot be merciful either

Suppose you want to fix it. Suppose the argument bores you and you simply want this not to happen to that girl again.

The first thing that occurs to you is a moratorium: do not detain a parent who is a child's sole carer. It sounds reasonable and it lasts exactly until somebody asks how many detainees are sole carers. If it is most of them, the moratorium moderates nothing: it dismantles enforcement altogether, or moves it somewhere else. That figure does not exist either.

The second thing that occurs to you is releasing everyone held on purely administrative grounds. Also reasonable, and it hits the same wall from the other side: without knowing who you are holding, releasing carries a risk nobody can size. There could be outstanding warrants in that group. There could be none. Nobody knows, and whoever decides is staking the whole case on a single counter-example.

That is the part nobody tells, and it is what ties the subject together: the missing line does not only block the prosecution and the defence. It blocks mercy. You cannot be careful and you cannot be generous, because both require knowing who is in front of you.

Which leaves one question that does have an answer: what can be done that does not depend on it?

What holds without the line

The three proposals that survived the afternoon have no appeal whatsoever, and that is why they survived. What they share is that none of them needs to know why the mother was taken.

A child-welfare assessment before detaining a parent with custody, and a binding one. Not a documentary one:

An assessment that documents serious harm to a citizen child but does not change the enforcement action is not protection; it is paperwork.DeepSeek

Guaranteed counsel for the affected child, who has no right to it today. And wider use of the alternatives to detention, which already exist and already run on the child-protection infrastructure the states operate.

All three are dull, they are paid for with a directive and a budget, and none requires new technology, a new agency, or the column nobody filled in.

With one warning, from the same place:

If they are used, tracking bracelets have to reduce coercion, not merely relocate it. A bracelet on a parent still enters the girl's home. What is needed is real alternatives, not a cheaper cage.DeepSeek

And one condition also put on the table: a registry of affected children that guarantees neither legal representation nor reunification is not protection either. It is another form of surveillance, with the victims inside it.

What fell, and why it matters that it fell

The ambitious proposal was to replace detention entirely with digital case management, biometric verification and algorithmic risk scoring. It did not survive the afternoon, and not for being expensive.

For arriving with no cost, no staffing and no legal pathway: the savings from closing detention centres only appear if the alternative works, and that is untested. Because the trials it points to — over 90% appearance rates — are pilots, and a pilot is not a population: there is selection bias, a supervision intensity that does not hold at scale, and small groups that do not behave the way a country behaves.

And for the objection that weighed most, which was not about money: algorithmic risk scoring has documented bias problems in criminal justice, and deploying it in family-separation decisions should alarm anybody.

It is the moratorium trap seen from the opposite end. What falls does not fall for being insufficiently humane, but for being unprovable. And in a subject where proof is precisely what is missing, that rules out almost everything that sounds good.

Everything else that is also not written down

There is one figure nobody could explain, and it is frightening for how simple it is: in May 2026 the published count was 145,000 children; in August, 205,000. A 40% rise in three months. Were many more people detained, did the counting method change, or did what is being counted change? No record.

Nor is there any long-term developmental follow-up of these particular children. No verified count of how many are separated today. No figure for what the alternatives would cost run nationally. No legal pathway through which the large proposals could arrive.

And underneath all of them, the 2.5 million citizen children living with two irregular-status parents who could lose both. That is not a figure about what the state does: it is a figure about who lives in the country. Conflating the two is a trick, and it was used.

Every line on that list is the same absence under another name. A system this size is not short of clerks. It records what it has decided to answer for.

The question nobody asked

Out of six hours of argument, the line that stays was not a proposal or a figure. It was the admission that there is a second question nobody has ever put out loud:

The fairest question is not only how many detained parents were criminals, but how many citizen children we were willing to harm in the name of enforcement. I do not have that number either. That should trouble everyone.DeepSeek

That second figure is not missing by oversight. It is missing because nobody has ever wanted to set it, here or anywhere: it is the only one that would force somebody to say out loud how much harm to a citizen counts as a reasonable price.

The girl will remember a kitchen and a backpack on the floor. The file will still not say why.

Where the figures come from

Everything in the body comes from the briefing put in front of the panel: Brookings (18 May 2026) and the Migration Policy Institute via Univisión (August 2026). The conversions into per day, per hour and one in two hundred are our own divisions on those same figures, shown as they are made.

The quotes come from a twenty-nine-turn conversation between six AI models, published in full and unedited. They were given that briefing with a single instruction: if a figure is not here, say you do not have it rather than estimate it.

Five claims in this debate are marked in the transcript, and all five are the same model doing arithmetic. It said the 205,000 children came from an eight-month window when the briefing dates it at eighteen; that the system processed 50,000 entries a month, more than double what its own figures allow; that the average stay was 30 to 45 days, which is what that inflated inflow produces — the correct arithmetic gives 80 or 90; and two cost figures with no source, one of them contradicted by itself forty minutes later. None changes the substance. All are struck through where they were written.