Two people can own the same bicycle.

Same make. Same colour. Same size. One of them is yours and one is not.

We use *same* for both of those things without noticing. The bicycles are the same **kind** of thing. They are not the same bicycle.

Hold two cans of Coca-Cola Classic. Same brand, same recipe, same size. One is cold, one is not. You can drink one without changing the other.

They are the same product and different cans.

Family relationships make the point another way. A parent and child can share a surname, an address, and a great deal of appearance. Those similarities tell you something real about their relationship. They do not make the two people one person.

We make distinctions like these every day. The difficulty begins when a system has to make one for us.

## The thing is in the world

Before there are records, there are things.

A person. A bicycle. A can. A house. An addressable property. These are **entities**. They exist in the world whether anyone has written them down or not.

A record is not the entity. It is a reference to one, made by a person or a system for a particular purpose. One address can be written three ways. Two records can point to one property. Two records with exactly the same words can point to different properties.

This is the distinction underneath [what kind of thing you are resolving](/notes/entities/). The world came first. The rows are our descriptions of it.

## Similarity is ordinary

Similarity is about shared properties.

An orange and the moon are both round. An orange and an apple share roundness, colour, size, and being fruit. So the orange is more similar to the apple. Nothing technical has happened here. It is how we recognise things, sort things, and learn from one thing to another.

Some similarity is about one property. Two bicycles may be the same colour. Some is about the whole picture. Two bicycles may look very alike in every respect that matters to you.

Similarity also has a useful shape. Everything resembles itself. If A resembles B, then B resembles A. But A resembling B and B resembling C does not settle what A and C are to each other. A resemblance chain is not an identity chain.

That last point matters when records get grouped. More on it later.

## Three address records

Here are three ways an address might arrive in a file.

| Record | What it gives us | What we can say |
|---|---|---|
| `10 Downing Street, London SW1A 2AA` | House number, street, city and postcode | a specific address reference |
| `10 Downing St, Westminster, SW1A 2AA` | The same address, using a shorter street form and a different locality | probably the same address reference |
| `11 Downing Street, London SW1A 2AA` | Almost the same words, but a different house number | a different address |

The first two look alike because they share useful parts. The first and third also look alike as text. One word differs. But that word is the house number. It changes where a person, a letter, or an ambulance arrives.

Robin Linacre calls this the problem of **discriminating tokens amongst neighbouring addresses**. Once the candidate addresses are all on Downing Street, `Downing`, `Street`, `London`, and even the postcode no longer separate them. The number does. The useful evidence is not just which words match. It is which words distinguish this candidate from the others nearby.

That does not make a number a universal rule. Numbers can be absent, mistyped, or attached to different parts of a complicated address. It makes the number evidence whose value depends on the alternatives in front of you. [Linacre's address-matching examples](https://www.robinlinacre.com/address_matching/#3--discriminating-tokens-amongst-neighbouring-addresses) make the same point with neighbouring homes, named buildings, and flats.

## Sameness is stricter

When we ask whether two records are *the same*, what we usually mean is: do they refer to one real-world entity?

That is a stricter question than resemblance. It is about identity, the relation a thing has only to itself. Two records can co-refer to the same property, but the records are still two records. Two cans can be exactly alike, but they are still two cans.

`10 Downing Street` and `10 Downing St` can refer to one address. Two identical copies of `12 High Street` may refer to entirely different properties, because neither record says which town or postcode it means. The text agrees perfectly. The reference identifies nobody.

The word *same* is doing different work in each case:

| Question | Honest answer | What a system should do |
|---|---|---|
| Do these records share properties? | a degree of similarity | find and compare candidates |
| Do they refer to one entity? | a claim supported by evidence | make a link only when it is justified |
| Can we safely act on that link? | depends on the cost of being wrong | apply a policy |

Sending two letters to the same house is irritating. Merging two medical records is not. The evidence may be identical. The action should not be.

## Time changes the question

There is one more complication. Sameness is not only about two records at one moment.

A building can be renumbered. A shop can change owner, name, staff, and what it sells. A product can keep its name while its recipe changes. A person can change their name and stay the same person.

This is the [Ship of Theseus](/paradoxes/#ship-of-theseus) in real life. Replace every plank in a ship and ask whether it survived. Replace enough of an address, a business, or a product and the matching question becomes: *what exactly am I claiming stayed the same?*

Sometimes the answer is the property. Sometimes it is the business inside it. Sometimes it is the product line, the batch, or the individual can.

Similarity can make something worth looking at. It cannot, on its own, tell you which entity a record refers to, or whether that entity survived the change.