A rant.

I came to entity resolution through geospatial engineering. Places are where the whole problem started for me. I have been keeping [reading notes on spatial identity](https://github.com/denironyx/systems-that-decide-what-matters/tree/main/04-spatial-identity) for almost six months now, trying to understand why a point on a map is so often not enough.

"Meet me at Waterloo" is a complete sentence to a Londoner. To a system, it is the beginning of several questions. The railway station? The Underground station? The bridge? The road? The wider area?

Open Google Maps and it will give you useful answers. A pin. A route. A label. Perhaps a place ID. But those are not Waterloo itself. They are references made by a system to parts of a real world that was there before the map and will remain when the app changes.

That is the same distinction behind [entities](/notes/entities/). An entity exists in the real world. An information system holds references to it. A place is an entity in this sense. Google Maps does not contain Waterloo. It contains names, geometry, identifiers, categories, images and claims about Waterloo.

The trouble begins when we mistake those references for the thing.

## Space is what the GPS sees

Edward Relph drew the distinction in 1976, in *Place and Placelessness*. Yi-Fu Tuan sharpened it a year later.

**Space** is abstract. Coordinates, extents, distance.

**Place** is dynamic, complex, it is what happens when people claim a piece of space by naming it, using it, and staying long enough for it to accumulate meaning. Relph's line is that engineers create localities, but time is needed to create place.

A latitude and longitude is not space itself. It is a reference to space, a measurement that a system can store. A name, a boundary, a Google Maps pin, an address and a place ID are references too. They may point well. They are still not the place.

Waterloo will remain Waterloo if a mapping service moves its pin, changes a label, or disappears. The station, bridge, road and neighbourhood are distinct real-world things, with relationships between them. A system has to decide which one a reference means.

This is not romantic. It is a claim about which references are stable enough for the decision in front of you. Coordinates for one building can disagree because one person measured at the gate, another at the main entrance, and another from a desk with a map.

The name, the type, and the named things that contain it may help decide what is meant. They can drift too.

Space is precise and unreliable. Place is fuzzy and durable. Any system has to hold both. But, the question is how?

Take *Elephant and Castle* in London. In conversation, it can mean the station, the junction, the former shopping centre, or the wider district. A coordinate can name one of those points exactly. It cannot decide which one a person meant without the rest of the reference.

## What a place gathers

Say *Manhattan* and a great deal comes with the word. An island. A borough. Neighbourhoods. Wealth, Streets, buildings, institutions, stories, and a changing idea of what the name stands for. *Brooklyn* works the same way. The place is physical, but what it means has been made over time.

The White House makes the distinction even interesting. It can mean a building, a residence, an office, or an administration speaking through the building. Those are not the same entity. They are related through place, language and history.

A coordinate can help you find the building. It cannot tell you which of those meanings a person intended. A place carries more than position. It carries the work people have done there, the relationships formed around it, and the references that have survived long enough to become shared knowledge.

There is a human side to this too. One of the first questions we ask somebody new is, "Where are you from?" The answer can be a country, a city, a neighbourhood, or the place that raised them. It changes with who is asking and what they are trying to know. Add the word *really*, and the question can become a refusal to accept the first answer.

This is one thread I want to take up in a future post, *Place identity*, on how the places we live shape how we understand ourselves and how other people try to understand us.

## The address was never about finding your way

Deirdre Mask's *The Address Book* has a reversal every identity person should sit with. Street addresses were not invented so you could find places. They were invented so the state could find **you**.

House numbering arrives in Europe in the 1700s as an instrument of taxation, conscription and policing.

And here is the detail that connects it to everything else. Mask reports that in 1700s England, ninety percent of men shared one of just eight first names. John, William, Henry and their five brothers had collided so completely that a name could not identify anyone.

States had tried fixing names before. Hereditary surnames, parish registers. The collisions kept coming.

House numbering attacked it from the other side. Pin each person to a numbered place, and an ambiguous name becomes resolvable.

That is the moment the address became an identity attribute. It was adopted because the name collision problem had already broken identification centuries earlier. Frequency-weighted matching and the postal address are two answers to the same problem, three hundred years apart.

## An address is not a place

An address is an agreement about how to find a street, building, entrance, or parcel of land. It is a very useful reference. It is not the place itself.

One address can contain many entities. A supermarket, the coffee shop inside it, and the pharmacy beside it can all receive post through the same address. One entity can have several addresses. A hospital has a campus, departments, entrances, and sometimes a postal address that points only to an administrative office.

Then there are places that people use and recognise without an address at all. The junction by the central mosque. The football pitch by the old market. They are not failed addresses. They are place references with a different structure.

This matters for matching. An address is often strong evidence that two references may concern the same place. It is rarely enough to decide it. You still need to know what kind of thing sits there, which entrance or unit is meant, and whether the reference was true at the time it was recorded.

The mistake is treating an address as an identity. It is one way a system reaches toward an identity, and it works only where somebody has done the long social work of assigning, maintaining, and sharing it.

## So how do you give a place an identity?

This is the question I actually want to think about, and there are only two families of answer.

**Derive it.** Compute an identifier from the attributes you already have. Same inputs, same key, and anyone can recompute it without asking permission.

**Assign it.** Mint an identifier, write it in a register, and maintain that register as the world changes.

These are the two main families of answer, and the choice determines what a system can and cannot do. It is worth looking at the best current example of each.

### Placekey derives

[Placekey](https://www.placekey.io/) is a free, open identifier for physical places, backed by a few hundred organisations including Esri and CARTO. Its structure is the interesting part.

It has two halves. The **What** encodes the address and, optionally, the point of interest at that address, so a coffee shop inside a supermarket is distinguishable from the supermarket. The **Where** is a hexagon from Uber's H3 grid, at resolution 10, roughly fifteen thousand square metres.

The appeal is real. No registry to join, no permission to ask, no central authority to trust. Two organisations can compute the same key for the same place and join their data without ever talking to each other.

That is a genuinely good property, and it is why the thing has spread.

But look at what it derives *from*. The What half is built from a postal address. Placekey's own documentation says it works for places with postal addresses, and full support currently covers the United States, Canada and the Netherlands.

Hold that thought.

### Overture assigns

The [Global Entity Reference System](https://docs.overturemaps.org/gers/) takes the other road. Overture mints an ID for an entity and commits to keeping it stable across releases.

To make that work it runs infrastructure: a monthly reference map of the current canonical state, a registry of every ID ever published, a changelog of what changed between releases, and bridge files connecting IDs back to the source records they came from.

That is a lot of machinery, and the machinery is the point. Because the identifier is not computed from the attributes, the attributes can change without the identity changing. A shop can be renamed, remeasured and recategorised, and it remains the same entity.

The cost is that somebody has to run it, forever, and you have to trust them.

### UPRN assigns, in one country

Britain has a strong version of this idea. A [Unique Property Reference Number](https://www.ordnancesurvey.co.uk/products/os-open-uprn), or UPRN, is a persistent numeric identifier for an addressable location in Great Britain. It is assigned and maintained through a property's life, from planning through demolition, even when its address or use changes.

This is what good reference infrastructure looks like. A council, a utility and a health service can use the same UPRN to join their records without treating their differently formatted addresses as different places. The identifier is openly available. The work of maintaining it is not magic. Local authorities, Ordnance Survey and GeoPlace do it.

But a UPRN still refers to an addressable location, not to everything that happens there. One building can contain a dozen businesses. A university can span many UPRNs. The property reference is strong evidence in a resolution decision. It is not the cafe, the company, the clinic, or the people inside it.

### The trade-off underneath

Put the two side by side and a general shape appears.

| | Derive | Assign |
|---|---|---|
| Identity comes from | the attributes | a decision, recorded |
| Needs a central party | no | yes, permanently |
| When attributes change | the key changes | the identity survives |
| When attributes are missing | no key at all | fine |

Neither is a bad design. They are answers to different fears. Derivation is afraid of dependence on an authority. Assignment is afraid of the world changing underneath a key.

And the deeper point is that this is not really a geospatial question. Identity is never in the data. So either you compute a proxy from data and inherit every weakness of your inputs, or you appoint someone to decide and inherit their politics.

Places just make the trade unusually visible.

## What derivation costs

Back to that held thought. If your identifier is derived from a postal address, then people without postal addresses do not get identifiers.

The exclusion is not marginal. More than a billion people live in slums or informal settlements, where recognised addresses are often absent or do not describe how people actually find a home.

So the elegant, permissionless, no authority required design fails precisely where an address was never available to derive from. Not by oversight. By the choice of what to derive from.

This is not an argument against Placekey, which does what it says on the tin in the countries it covers. It is an argument about what "universal" means when the input is an artefact of state administration that most of the world never received.

Grid codes are the obvious response, and they are genuinely useful. Plus Codes have been painted on doorways in Kolkata by a non-profit, giving people something a delivery driver can use. what3words divides the earth into three-metre squares and replaces each square with three words. Both can make an otherwise hard to describe location easier to communicate.

## A code can find the square, not name the place

Google's Plus Codes are more properly called Open Location Codes. A full code is derived from latitude and longitude, and anyone can encode or decode it offline. A Plus Code is not locked inside Google Maps in the way a Netflix account number is locked inside Netflix.

what3words takes another route. It gives every three-metre square three words. It works in the limited and important sense that it gives someone a precise way to communicate a location. But the grid, word lists and conversion service are held by one private company, under its licence. A three-word address can travel between people, but its meaning still depends on that company continuing to resolve it.

Mask's warning maps cleanly onto Relph. A grid code is pure space. No community has to speak it. No history attaches to it. Even where a delivery or emergency service accepts it, the code can live in an app rather than in neighbours' heads. It solves a real problem. It does not create place.

## The seam where the digital becomes physical

Mask's story is a state reaching a subject. Change the actor and the argument walks into the present. Now it is a merchant trying to reach a customer, and it runs on the same field.

Consider what you are to a shop you have never visited. An account, a device fingerprint, a payment token, a session, and a history of things you looked at and did not buy. Every one is an identifier that exists inside a system. None has a location, and none needs one. The whole apparatus works until something has to arrive. Then it hands off to a rider deciding whether this is the right gate.

That handoff deserves more attention than it gets. The order is bits. The payment is bits. Fraud scoring, the recommendation, the confirmation email, the tracking page with its moving dot, all bits. Each can be retried, cached, replicated or rolled back. The last hundred metres cannot. It happens once in the street, and nothing upstream can compensate for getting it wrong.

I would resist calling that meeting a *point*. The geometry is precisely the thing that fails. Call it a **seam**. A seam is where two materials made separately have to be joined, and it is where things tear. The delivery does not fail because a coordinate was wrong in the fourth decimal place. It fails because two parties do not share a representation of the same place. The customer has one, "the blue gate after the mosque, ask for Mrs. Grace". The system has another, a pin dropped onto a roof from a satellite image. The rider has to reconcile them with the engine running.

Precision is not the missing ingredient. A shared, checkable representation is. That is the same claim this note makes about matching, arriving from the other direction.

The seam runs both ways. **Digital to physical** is a parcel, a dispatched ambulance, a meter installation, or a ballot box. It fails loudly. The package does not arrive, somebody calls, a refund is issued, and a complaint is logged. **Physical to digital** is proof of address for a bank account, a census enumeration, a facility register, or a tax roll. It fails silently. You are simply not in the file, and nobody phones to say you were left out.

Commerce has noticed the seam and is fixing it independently. Riders keep notes. Apps save pins. Platforms learn internal place maps from successful deliveries, so the eleventh parcel may arrive more easily than the first. That is a place representation, learned from physical outcomes rather than declared by a surveyor. It is also owned, unshared and non-portable.

Your address becomes solved inside one logistics stack and stays unsolved everywhere else. The ambulance, meter reader and next platform may each have to rediscover the blue gate from scratch.

One correction before the frame gets too tidy. The address is not the only bridge, and in much of the world it is not the primary one. A phone number does an enormous amount of this work. One-time passwords, mobile money, and most tellingly the call a rider makes from the junction. Delivery in Lagos or Nairobi is often a two-stage protocol. The address or landmark gets you to the neighbourhood. A voice call gets you to the door.

So the narrower claim is this. Place is the bridge for anything that has to arrive. The phone call is the fallback protocol we use when the place representation was not good enough to carry the handoff alone. Counting how often that fallback fires would be a decent measure of how badly a place is represented.

## Places drift, so the identity has to date

Relph's deepest observation is temporal. Time makes place, and time keeps remaking it.

Places do not change at the same rate. A restaurant can close, reopen, and change hands in a year. A hospital or university may hold its name and role for decades.

There is also category drift, which is worse because it is quiet. A business stays at the same address with the same pin, and changes what it does. The coordinate stays right while everything else goes wrong.

So any single refresh cadence across a place database is wrong for some large part of it. The identifier can stay stable. The resolution claim that attaches it to a reference needs a date.

The same building can persist while its name, use, tenant, or recorded entrance changes. A system that returns a bare yes has thrown away the only honest part of the answer, which is *when*.

## Distance proposes

The matching work says the same thing more plainly. In Foursquare's public location-matching competition, the strongest approaches combined proximity with names, categories and other comparison features. Distance generated candidates. It did not decide the match.

In a dense area, two coffee shops thirty metres apart are two entities. One hospital measured twice at two kilometres can be one. Space proposes. Place decides.

## The sensitive half

Everything above treats places as public. A clinic, a junction, a charging point. Resolving them well helps everyone.

Make the place a home and the same machinery turns dangerous.

Golle and Partridge showed in 2009 that the pair of your home and work locations, even coarsened, narrows most people to a handful of candidates. De Montjoye and colleagues showed in 2013 that four rough spatio-temporal points uniquely identify ninety-five percent of people in a mobility dataset of one and a half million.

Location is not an attribute *about* you. Past a certain precision it simply *is* you.

Which means a place identifier for a residence is a person identifier wearing a different coat, and should be governed like one.

## A landmark is not a failed address

"Behind the filling station, at the junction" gets used as an example of an address that breaks parsers. I want to read it the other way.

It is a rich place description. A landmark anchor, an implied containment, a spatial relation, and a verification community. Thousands of people can confirm it, correct it, and navigate by it.

It is *more* place-like than "Flat 2, 12 High Street", not less. What it lacks is not information. What it lacks is a system willing to represent it.

And the cost of that is not inconvenience. The order gets placed, the payment clears, the digital half runs perfectly, and then the last hundred metres has to be improvised over the phone by two people with no shared way to name the same gate.

Whatever cannot be improvised over the phone does not arrive. Not the parcel, and not the ambulance, the meter reader, the enumerator, or the ballot.

## Lagos knows before the database does

In Lagos, "Computer Village, Ikeja" can be enough to begin. It names a real commercial area that people know. It is not enough to finish a delivery, because the next question is which shop, which building, which entrance, and who is there today.

That does not make the reference poor. It makes it honest about its level. A map pin might give a driver somewhere to start. A landmark, a caller, a shop name and the current route finish the job. Each reference carries part of the answer.

The mistake is asking one reference to do all the work. A location code can name a square. A UPRN can name an addressable property. A business listing can name a shop. None of them is the whole place, and all of them need dates, evidence and context before a system acts on them.

## Where I run out

I do not know how you build the assign-side registry for places that no state ever registered.

Derivation fails there because there is no address to derive from. Assignment fails there because assignment needs an institution, and the absence of an institution is precisely the condition being described. Grid codes fill the gap functionally and create nothing that behaves like a place.

The one thing I am fairly sure of is that the delivery platforms are slowly solving it, per-company, out of successful drop-offs. Each of them is learning an excellent representation of your gate and each of them owns it privately.

Which means your address gets solved inside one logistics stack and stays unsolved everywhere else. The ambulance still cannot find you.

That looks less like a technical problem than a question about who is allowed to hold the answer.

## References

- Edward Relph, *Place and Placelessness*, Pion, 1976.
- Yi-Fu Tuan, *Space and Place: The Perspective of Experience*, University of Minnesota Press, 1977.
- Doreen Massey, "A Global Sense of Place," *Marxism Today*, 1991. The relational rejoinder.
- Deirdre Mask, *The Address Book*, St. Martin's Press, 2020.
- [Placekey](https://www.placekey.io/faq), on the What/Where structure and H3 resolution 10.
- [Ordnance Survey, OS Open UPRN](https://www.ordnancesurvey.co.uk/products/os-open-uprn), on persistent property references for addressable locations in Great Britain.
- [Overture Maps, Global Entity Reference System](https://docs.overturemaps.org/gers/).
- [Foursquare Location Matching](https://www.kaggle.com/competitions/foursquare-location-matching), Kaggle, 2022.
- [Foursquare OS Places](https://opensource.foursquare.com/os-places/), open sourced 2024, Apache 2.0.
- [Open Location Code](https://github.com/google/open-location-code), the open specification and reference implementations behind Plus Codes.
- [Plus Codes, addressing for Kolkata citizens](https://maps.google.com/pluscodes/case-studies/addressing-for-kolkata-citizens/).
- [what3words terms and conditions](https://what3words.com/terms), on the company-controlled data and licence behind three-word addresses.
- [World Bank, urban development](https://www.worldbank.org/ext/en/topic/urban-development), on the global scale of informal settlements.
- Golle and Partridge, "On the Anonymity of Home/Work Location Pairs," *Pervasive Computing*, 2009.
- de Montjoye et al., "Unique in the Crowd," *Scientific Reports* 3:1376, 2013.
- Peter Christen, *Data Matching*, Springer, 2012.
- Christen, Ranbaduge and Schnell, *Linking Sensitive Data*, Springer, 2020.
- My reading notes on [spatial identity](https://github.com/denironyx/systems-that-decide-what-matters/tree/main/04-spatial-identity).