In 2019, scoring a rental applicant on anything past income and a credit file took a vendor, a contract, a minimum seat count and a budget line. Four property groups in the city could carry that. Everyone else read the references and formed an impression.
By 2026 the model runs on a laptop, and the woman applying for the two-bedroom is scored four times before anyone meets her. The letting agency scores her ability to pay. Her name goes into the block’s risk pool at the insurer underwriting it. The management company estimates how much of its staff time she’ll take. And the landlord, who has never scored anyone in his life, runs something a friend sent him. None of the four knows the others did it.
Each tenant score was allowed. The total wasn’t.
View source
%% title: Four grants, one tenant
%% caption: Each arrow is a relationship the sequence's tests can read: one institution, one grant, one exercise. The count of arrows is a fact about her, and no test on any single arrow reaches it.
flowchart TD
classDef system stroke-width:1.2px;
classDef actor stroke-width:1.6px;
classDef status stroke-dasharray:4 3;
subgraph Scorers ["Four independent scorers"]
direction TB
A["Letting agency (ability to pay)"]:::system
B["Insurer (block risk)"]:::system
C["Management co. (staff time)"]:::system
D["Landlord (own model)"]:::system
A ~~~ B ~~~ C ~~~ D
end
T((The applicant)):::actor
E["Jurisdictional density: 4<br/>(read by no test above)"]:::status
Scorers --> T
T -.-> E
From four firms to every landlord with a laptop
The old arrangement kept the capability inside four firms. The new one puts it within reach of every agency, every underwriter, every managing company and every landlord with a laptop. Measured the way concentration is normally measured, this is a clear improvement. The vendor’s pricing power is gone. So is the lock-in. Four companies no longer decide who in the city gets scored and who gets read by a human.
Yet her exposure went from one score to four. Both sentences are true at once, and tests that take one institution at a time can see only the first of them.
Run the tests on all four and they come back clean. The agency assesses whether an applicant can pay the rent, which is what a letting agency was authorized to do before software existed. It treats her the way its grant allowed, on the features that justified the grant. Its question, decision and consequence are unchanged, and only the estimate is sharper. The insurer prices risk, the managing company forecasts its own labor, and the landlord chooses a tenant, each inside its own grant.
Across four grants and four exercises inside them, nothing reaches past what any one of these actors was always entitled to decide. The tests take one institution and one person and ask whether that grant still covers what’s being done, so each reads one arrow and has nowhere to put the other three.
Legitimacy does not sum
The four scores don’t fail independently, and one bad entry in her file shows how: an eviction case filed against her six years ago and dismissed. Tenant screening reports carry records like that, and the Consumer Financial Protection Bureau found that an eviction record tends to end in a denial whatever the outcome of the case.1 The agency’s model reads the entry, and if the insurer and the landlord pull from the same screening bureau, their models read it too. Kathleen Creel and Deborah Hellman argued that when algorithms built on the same data decide in many contexts, someone can be shut out of a whole range of opportunities. An error a person reading her references might have asked about once is counted against her four times by parties that never compare notes.2
The next application runs through the same kinds of models with the same entry in the same place. If she finds the error and gets the agency to fix its copy, the other three still hold theirs. Contesting it means four procedures with four sets of forms and four clocks, and she has to know a score exists before she can contest it. Usually she learns of one, from a denial letter, and never of the rest.
Legitimacy does not travel: a grant covers what it covered and thins with every step past it. Density shows a second problem, because legitimacy does not add up either.
Jurisdictional density is the count of institutions that govern one person at the same time. It’s a fact about her that belongs to none of the four, and nobody can see it, her included.
Four adequate grants don’t make an adequate situation for her, because no grant was ever measured against the others. No legitimacy test runs across institutions to ask what they add up to at the point where they meet, and that is the blind spot density measures.
It needs no villain. Four parties produce it by each doing exactly what the tests require, and the dismissed case counted four times lives entirely in the space between their grants.
Who could answer for the count
The capability still shouldn’t have stayed with four firms. For years the landlord didn’t score anyone because of the budget, not because any rule said he may not, and a price is a bottleneck rather than a boundary. Concentration rationed her exposure as a side effect of protecting four companies’ margins. Anyone who wants that back is asking for the same scoring sold at a markup, reaching fewer people by accident.
Breaking a capability out of a few hands redistributes who holds it and says nothing about how many institutions end up governing any one person, so deconcentration by itself points nowhere in particular. Those two quantities were always different, and they’ve been treated as one because, for most of the history of the argument, concentration was where the cases were.
An individual right works one institution at a time, which is the shape of the problem rather than a solution to it. None of the parties produced the condition alone, and no rule makes them answer to each other or to her jointly.
The count needs feedback with standing at its own size. That means a way to make the hidden count visible, gather the people who carry it, and give that group standing because of the total rather than any one score. The closest existing forms gather many people against one institution’s rule. Density needs the reverse, several institutions answering together for one count, and no such body exists yet.
All four could produce their grants on demand. She could produce nothing, because what happened to her happened between them.
Notes
Consumer Financial Protection Bureau, Tenant Background Checks Market (November 2022), on outdated, incorrect and incomplete eviction records in screening reports.
Kathleen Creel and Deborah Hellman, “The Algorithmic Leviathan: Arbitrariness, Fairness, and Opportunity in Algorithmic Decision-Making Systems,” Canadian Journal of Philosophy 52, no. 1 (2022): 26–43, argued that an arbitrary algorithmic decision is rarely a wrong in itself, but that when the same algorithm, or algorithms built on the same data, decide in many contexts, a person can be shut out of a whole range of opportunities. Rishi Bommasani and colleagues, “Picking on the Same Person: Does Algorithmic Monoculture Lead to Outcome Homogenization?,” Advances in Neural Information Processing Systems 35 (2022): 3663–78, found that decision-makers who share training data more often give the same individuals a negative outcome.