The Sorting Machine
The same machine an earlier essay watched sort children, now sorting homes, operating with constitutional force under existential stress, with a language model bolted on top.
The village
In a de-occupied village in Kharkiv oblast, a woman in her seventies stands in front of what used to be her house. She has owned it since 1986. The deed is a paper booklet in a drawer that no longer has a wall behind it. To be compensated for the loss, she will need to open an app called Diia on a smartphone she does not own, and prove she owned a house that the state’s digital property registry has no record of. Ukraine has built the most advanced wartime digital state on earth to help her. By its own design, it cannot see her.
This is not a story about a bug. Everything in that paragraph is working as specified. The app verifies ownership against a registry, the way it is supposed to. The registry contains what was digitized, the way it is supposed to. The woman does not appear, because she was never entered. The machine is not failing her. It is classifying her, correctly, as a person it has no record of, and distributing the consequence that classification carries.
An earlier essay in this series described this machine in American special-education offices, where a classification system decided which children received support and which received labels, and where the label, once assigned, became the thing the institution then confirmed. Move the machine to a country under invasion and the category it assigns is no longer a school placement. It is a home. The stakes change completely. The machinery does not change at all.
The frame
Geoffrey Bowker and Susan Leigh Star gave us the load-bearing idea.1 The most powerful classification systems are the ones nobody experiences as classification at all. They sink into the background, become infrastructure, and from there they sort the world while remaining invisible to the people being sorted. Star catalogued the properties that make a system infrastructural,2 and three of them do almost all the work here: a system inherits the limits of the base it was built on, it produces a residual category for everything that does not fit, and it stays transparent to those it fits while becoming a wall to those it does not.
A wartime state has to classify at speed and at scale. Which damage, whose property, which category, how much, paid to whom. There is no time to do it by hand and there is no peacetime registry deep enough to lean on. So the classification gets built fast, under fire, and then it acquires the force of law: it decides who is restored and who is not. That is the same place the companion piece in this series, Unratified, located the identity layer of digital public infrastructure. A system that carries constitutional force without ever having had a constitutional moment. eRecovery is that argument pointed at property instead of identity.
“Classification systems provide both a warrant and a tool for forgetting… the moral questions are buried in technical decisions.”
Bowker & Star, Sorting Things Out (1999)
The machine
Here is the machinery, concretely. In May 2023 Ukraine launched eRecovery through Diia: a citizen applies for compensation for damaged or destroyed property directly in the app.3 Alongside it, the Council of Europe established the Register of Damage for Ukraine, which sorts claims into more than forty coded categories, grouped A for individuals, B for the state, and C for legal entities, opened in stages, with claims filed through Diia.8 Damage to a home is a category. It has a code. The category scheme is not a metaphor for Bowker and Star’s argument. It is a literal instance of it.
To pay you, the machine first has to confirm you own what you say you owned. It does this against the State Register of Property Rights. And it operates at the scale of a whole society, not a pilot: more than 14 million people use the Diia ecosystem, by the ministry’s own count roughly every fourth citizen,9 and a 2025 UNDP and KIIS study found that around 59 percent of Ukrainians use government e-services.5 This is not a fringe portal. It is the front door of the state. Which is exactly why the question of who cannot get through it is not a footnote.
Run a claim
Before the argument, the machine. Configure a claimant and watch where the claim stops. Three gates, each one a failure mode that public reporting has already documented. Notice that none of the gates ask the question you would think a compensation system exists to ask, which is whether the house is gone.
Configure a claimant. Watch where the claim stops. Each gate is a documented failure mode, sourced below.
The claimant opens Diia, signs in, and starts an eRecovery application.
Diia verifies ownership against the registry and the property is found.
The damage maps onto an open compensation category and the claim proceeds to adjudication.
Every gate passed. This is the claim the machine was built to process, and it processes it fast.
This claimant has a smartphone and a Diia account, owns a flat registered after 2013, and lost it to a strike that fits an open damage category.
Three predictions Bowker and Star already made
The failures the funnel produces are not surprises. They are the three infrastructure properties, showing up on schedule, with names and numbers attached.
Star’s rule: infrastructure inherits the strengths and the limits of the base it was built on. It does not start clean. The State Register of Property Rights is about 40 percent complete, and is mostly populated with property registered after 2013, when digitization began.3 So a claim’s success depends less on whether the house was destroyed than on whether the ownership was digitized a decade before the war. The machine inherited a pre-war gap and turned it into a wartime exclusion.
Every classification makes an “other” box, and the other box has a population. The self-built home with no formal title. The house inherited and never re-registered. Damage in a category not yet opened. These claimants are not so much rejected as unrepresentable: there is no field for them. Bowker and Star called this the residual category, and warned that it is where a classification quietly stores the people it was not designed to hold. Here it holds people whose houses are gone.
Infrastructure is transparent to those it fits and a wall to those it does not. 38 percent of Ukrainians over 70 do not use the internet,4 which means the citizens most likely to hold a Soviet-era paper deed are the citizens least able to file through an app. The exclusions compound on the same person: offline, and pre-2013, and most likely to live where the fighting was. Three independent gates, one body standing behind all of them.
The agent on top
In September 2025 Ukraine added the part that makes this the wartime edition and not just a wartime example. It launched Diia.AI, billed as the world’s first national AI assistant that delivers government services rather than merely advising on them. By the Kyiv Independent’s account it was tested by 100,000 users in its first 24 hours and passed a million within a week.6 A citizen can now ask, in plain language, how to get compensation for a destroyed home, and an LLM agent will try to do it for them.
The architecture is the interesting part, and it is genuinely well built. According to Kitsoft’s chief executive Oleksandr Iefremov, the agent never touches the registries directly.7 A Model Context Protocol “defines the interface through which the AI agent can initiate a limited set of pre-approved actions.” A controlled service layer “prevents the AI agent direct access to state registries or service business logic.” Registry data reaches it through Trembita, Ukraine’s data-exchange bus, rather than by the model reaching in. This is exactly the safety design you would want from a state agent. It is also the place where the two halves of this research program meet: an agent architecture sitting on top of, and mediated through, a state classification system.
Diia.AI, top to bottom. Tap a layer. Read what it does, then read what it forecloses.
Defines the interface through which the agent can initiate a limited set of pre-approved actions.
“Find the house the registry does not contain” is not a pre-approved action. The protocol is the leash, and the leash is the point.
A language model that politely cannot find your house is, underneath the conversation, the same machine that could not find it before.
Read the stack and the load-bearing point arrives on its own. The agent can only take pre-approved actions, and “find the house the registry does not contain” is not one of them. It reads through a bus that can only pass what the registries hold. The controlled service layer that makes the agent safe to deploy is the same layer that guarantees it inherits, and cannot override, every exclusion below it. The 40 percent gap does not get smarter because there is a language model in front of it. The agent is courteous, fluent, fast, and structurally incapable of seeing the woman in the village, for the same reason the app was: she is not in the base.
What this is not
The honest concession, because the argument is worse without it. This is not a claim that Ukraine’s digital state is bad, and it is not “AI is bad.” The opposite, on both counts. eRecovery has moved real money to real people faster than any paper process could have managed, in the middle of an invasion, which is a genuine achievement and not a small one. The registry is being completed about as fast as a country under fire can complete a registry. Diia.AI is the most ambitious public-sector agent deployment anywhere, and its safety architecture is better than most of what the private sector ships.
The point is not that the builders were careless. It is the reverse. Classification infrastructure, when it works this well and this fast, is precisely when its exclusions stop looking like accidents and start looking like fate. A clumsy system fails visibly and gets fixed. A good one draws its line so cleanly that the line reads as the natural shape of the world. The better the machine, the more its residual category disappears from view, and the more it needs someone standing outside it to point and say: there are people on the other side of that, and their houses are also gone.
What it would take
The prescription is narrow, and it follows from the diagnosis rather than from a mood about technology.
A human-adjudicated channel that does not require the smartphone, the post-2013 deed, or the pre-approved action. When a system carries constitutional force, the fallback is constitutional-grade infrastructure too, not a phone number that rings in a basement. This is the argument Unratified makes in full: entrenched rights of exit are part of the design, not an apology for it.
Diia is rightly proud that its AI now handles a large share of support inquiries. That number measures the people the machine can serve. The number that matters for legitimacy is the one nobody publishes: how many claimants fall out at the identity gate, the registry gate, and the category gate, and who they are. A residual-category dashboard would make the wall visible to the people who built it.
For the pre-2013 owner, the missing registry entry is the entire barrier. Digitizing those deeds is not back-office hygiene. It is the difference between being compensated and being unrepresentable, and it should be resourced as such.
The return
Back to the village. The machine is not cruel to the woman standing in front of her rubble. It is correct. It will verify her neighbor’s 2019 apartment in minutes and pay out, and it will have nothing to offer her, politely, in fluent Ukrainian, through the world’s first national AI assistant. That is what classification infrastructure does at its best. It makes the line it draws feel like the natural order of things, right up until you are the one standing on the far side of it, in front of the place your house used to be.
The earlier essay watched the machine sort children. This one watched it sort homes, under fire, with an agent on top. The reference materials change. The categories change. The country changes. The machine is the same machine. It always was.
The same essay, two countries and two stakes. The Sorting Machine watches special-education placement sort children for compliance; the Wartime Edition watches Ukraine's eRecovery sort homes under fire. Same residual category, same inherited installed base, far higher stakes.
Classification as Infrastructure names the pattern in the abstract and inside AI systems. The Wartime Edition shows it converging in one live system: an LLM agent sitting on top of a state classification machine, mediated through MCP and the Trembita bus, inheriting every category the registry already enforces.
Two readings of digital public infrastructure that carries constitutional force without a ratifying moment. Unratified builds the identity layer in the reader's hands; the Wartime Edition points the same argument at property, where the category assigned is a home and the adjudicator is a wartime state.
Both apply Bowker and Star to a system that became constitutional without meaning to. The Invisible Architecture traces it in psychiatric diagnosis; the Wartime Edition traces it in Ukraine's damaged-property registry, where the residual category holds people whose houses are gone.
What the State Keeps reads the state as a classified inventory evaluated from above. The Wartime Edition reads one wartime state's compensation machine from below, from inside the residual category it cannot represent.
References
Every figure in this essay is drawn from public reporting and is cited here. Tap a superscript in the text to see the note inline, or expand a card below for the full source and a link.