Who is who

A name is not a person. Within one repository a spelling stands for one person; across sources, two entries are joined only where their records share an identifier such as an ORCID iD, or where the join was checked on a sample read by a person, and never on a shared name alone. A match a person confirmed is shown apart from one that is only likely.

Two people named Li Wang at two schools stay two people, since no identifier joins them; one person written Müller, Anna on one record and Mueller, A. on another is one person, since both records give the same ORCID iD.

Why a name is not a person

A thesis record names its people as its cataloguer typed them, and names are unreliable keys. One person appears with a full given name on one record and an initial on the next; family name first in one repository and last in another; with or without an accent, a hyphen or a second given name; transliterated from Cyrillic, Greek or Chinese in more than one way; with a title, a degree or a note in parentheses attached. And two people share a name more often than one expects: a common family name with a common initial can belong to several scholars in one field, sometimes in one department.

So the record holds to one rule. One person spelled two ways stays two people until a shared identifier, a check on a sample or a person's own reading says otherwise, and nothing joins people across sources but a shared identifier or a checked rule. Within one repository the rule cuts the other way: two people who share a spelling there share one entry, a limit set out below. A join that is wrong does more harm than a gap: it gives one scholar another's students, and every line drawn through them.

Within one source

Inside one repository, each name is put in a single form, one set of rules for every source: the text normalised, a note in parentheses and a title such as Professor, Dr. or Prof. Dr. taken off, in any of the languages the sources write them in; a name with no comma turned so the family name comes first, a block in capitals read as the family name wherever it stands; a full stop after a word at the end of a part taken off, an initial's kept; and the whole compared without regard to case. Each set of rules was checked on what it joins before it was used.

Within that source, one form is one person. That is what lets all of a supervisor's students at one university gather under one entry. It also has two limits, said plainly. Two people of the same spelling at one repository share one entry, since nothing in the record tells them apart. And a spelling that differs by an accent, a hyphen or an initial more is another person, even where a reader can see it is the same one. A value that holds several people no rule could split is kept whole and marked as several, rather than losing any of them.

Across sources: shared identifiers

Between sources, a name joins no one. What joins two entries is an identifier both records give, from a register that issues one to each person: an ORCID iD, which researchers hold for themselves; IdRef, the authority file of French universities; the GND, the authority file of the German National Library; a Wikidata item; an OpenAlex author record.

Where a repository's records give a person an identifier, it is kept on the person as the records state it. Where those records give two different values of one kind, the person is given neither, since which is theirs cannot be told. An identifier is only as good as the record that gives it: one repository attaches the same ORCID iD to every thesis of a name, so there it proves no more than the name does, and a join by it is held back where the name's own theses show it may be two people.

A join by identifier is a rule of its own and is checked like any other before it is used: a person whose records give an ORCID iD that Wikidata states for one item, the names agreeing, is joined to that item; one whose ORCID iD Wikidata holds for no one is joined to the OpenAlex author record that gives it. Joins by IdRef and by the GND are written and wait on their own checks: until they pass, those identifiers are kept and shown and join no one. A few joins were also read and confirmed by hand, one at a time.

Joins past the name

Many people share no identifier across their records: a person on Wikidata with no OpenAlex author stated for them, or one scholar whose papers OpenAlex has spread over several author records. For these the record tries rules that weigh evidence past the name, such as the years of a life, the places someone worked, a field, coauthors in common or a paper listed under both records. Three such joins have been built:

OpenAlex author records to the Wikidata people they belong to. Its check passed, and its joins are used.
One person on several OpenAlex author records: a stray record joined to the fuller one by a paper listed twice, coauthors in common, or the same place in the same years. Its first rules fell short of the bar and its second passed; the second's joins are used.
A Wikidata person with no OpenAlex author stated, to OpenAlex's records under their name. Two sets of rules were tried and neither reached the bar, so neither joins anyone by itself; a third is written and waits on its check.

Each is held to the same check. A sample of at least 100 of the pairs a join proposes is drawn at random, never from the people its rules were written on, and a person reads each pair against the sources and calls it yes, no or unsure. The share called yes, the unsure left out of it, must reach 0.90. A join that reaches the bar is used; one that falls short is not, and the result is kept all the same, since it says what the record declined to claim.

The check a join is held to: a sample of at least 100 of its pairs drawn at random, each read by a person as yes, no or unsure; the share right, the unsure left out, against 0.90; at or above it the join is used, below it it is not, or every pair it proposes is read by a person, a yes shown as a checked match and an unsure as a likely match, not confirmed.

A join that falls short can still serve as a list to read. That is how the third join entered the record: every pair the better of its two sets of rules proposed was read by a person, one by one. A pair read as the same person is shown as a checked match, one the reader was unsure of as a likely match, not confirmed, and a pair read as two people was left out.

How a match is shown

A person's page lists the registers the record links them to, and under each where the link was read: stated by Wikidata, given on their thesis record at a named repository, on their OpenAlex author record, or a match. A match says which it is, a checked match or a likely match, not confirmed; the likely one wears the dashed mark of a suggestion.

The difference carries through. The schema.org data of a person's page names another address as the same person only where a match was checked, never where it is only likely. An author record joined by a match is read for the person's papers alone: their coauthors, the authors they cite and the routes between them and others stay on Wikidata's own identifiers. Where no identifier finds an author record for someone, their card may offer a few OpenAlex records whose name agrees, each said to be unverified, never counted and never joined.

An address that holds

When two entries are joined, the person's address stays good. A page asked for at the address a person had before the join is sent on to the page they are under now, so a link someone saved or cited still arrives at the right person.

If two entries are one person, or one entry is two, say so from the person's page, which has a link to correct or remove the entry; the record changes only once the evidence, read as above, says it should. How the checks are scored is set out on measured, not asserted, and the kinds of evidence a link rests on on where every link comes from.