Academic Ancestry

About

Measured, not asserted

A way of reading a source is not taken on its writer's word. It is checked: a sample of what it read is compared with what the source itself shows, such as the thesis's own page or title page, and it passes where 0.90 of the sample or more agree. Every check is listed here as it was measured, the failures among them, and every count the site shows comes from a reading of the record, dated by the day it was made.

a sampleof what was readthe readingthe sourcecompareditem by itemthe bar, 0.90the sharethat agreedpassesat or abovefailsbelow, shown
One check, left to right: a sample of what a reading of a source gave; each item compared with what the source itself shows, such as the thesis's own page or title page, by a person, a program or a language model; the share that agreed held against the bar of 0.90; and the verdict, a pass at or above the bar and a fail below it, both kept and shown.

What a check is

A reading of a source can look right and be wrong. A field may hold a supervisor at one university and a department's head at another; a name may be written family name first in one repository and given name first in the next. So a new way of reading a source is not taken on its writer's word. It is checked.

A check draws a sample of what the reading gave: the supervisors it read from a hundred theses, say, or the pages it built addresses for. Each item is then compared with what the source itself shows, where a reader of the thesis would look: the thesis's own page at the repository, its title page or acknowledgements, the record as the repository served it, or, for one check, the advisors Wikidata states. Who makes the comparison differs, and each check on the board says who it was. Many are made by a program, which looks for each name or address on the source's own page, beside the label the page gives it. A few samples were read by a person, by hand. Others were read by a language model: the first checks of several formats' readers by a model with the raw records and the reader's rules open, no person reading the sample, and a few by Claude, the model the site works with, standing in for a person on the maintainer's word. Not one of these checks claims that a person read every sample. The share of the sample that agreed is the check's score, and the score is held against a bar fixed before the result was known.

The verdict is kept in a file of its own beside the code: what was checked, for which sources, on what day, how large the sample was, how many agreed, the score, the bar and whether it passed, with the words of whoever made the comparison. A check made since the site began keeping them this way also names the code it judged, by a fingerprint; once that code changes, the check no longer speaks for it until it is made again.

The bar

A check passes where at least 0.90 of its sample agreed: nine items in ten, read the way the source itself shows them. The bar is the same for every check, and it was set before any of them was made, so that no check could be passed by lowering it afterwards.

Two kinds of claim carry more. A link that joins two records of one person, and a supervisor read by a language model from a thesis's own text, are given no verdict at all on fewer than 100 items, and their score is reported with the lower end of its confidence interval beside it, so a reader can see how much of the margin is only the size of the sample. Some of the checks of readings below rest on fewer; each shows its sample beside its score, so the small ones can be weighed as small.

The checks as they stand

Every check of a reading of the sources is drawn here from the record as it stands, each a bar of its score against the bar of 0.90, the highest first. The chart's axis starts near the scores rather than at nothing, so that checks close to the bar can be told apart; its foot is labelled. A check that passed is drawn solid; one that failed is drawn hollow and marked. Under the chart each check is said in words: what it read and where, what it was compared with and by whom, how many of its sample agreed, and the day it was made.

0.801fails0.9050 of 51 pass the bar
Each of the 51 checks as a bar of its score, from 0.80 at the foot of the axis to 1 at its head, highest first, against the bar of 0.90: 50 passed, drawn solid, and 1 failed, drawn hollow and marked fails.

  1. The supervisors read, in the records of AUB ScholarWorks (American University of Beirut), compared with the source's own pages by a program

    passes: 106 of 106 agreed, 1.000 against 0.90, made 2026-10-19

  2. The role of the first person in a record, in the records of ScholarWorks at UMass Boston and CCU Digital Commons, compared with the source's own pages by a language model

    passes: 30 of 30 agreed, 1.000 against 0.90, made 2026-09-29

  3. The supervisors read, in the records of CORE: ELAR, Ural Federal University, compared with the source's own pages by a program

    passes: 108 of 108 agreed, 1.000 against 0.90, made 2026-10-20

  4. The supervisors read, in the records of Repositório UFMG; Repositório UFLA; Academica-e, UPNA; and Repositório UFPA, compared with the source's own pages by a program

    passes: 185 of 185 agreed, 1.000 against 0.90, made 2026-10-10

  5. The supervisors read, in the records of berkeley-math, compared with the advisors Wikidata states by a program

    passes: 100 of 100 agreed, 1.000 against 0.90, made 2026-10-20

  6. The supervisors read, in the records of berkeley-statistics, compared with the advisors Wikidata states by a program

    passes: 100 of 100 agreed, 1.000 against 0.90, made 2026-10-21

  7. The supervisors read, in the records of stanford-statistics, compared with the advisors Wikidata states by a program

    passes: 100 of 100 agreed, 1.000 against 0.90, made 2026-10-21

  8. The supervisors read, in the records of store, compared with the advisors Wikidata states by a program

    passes: 100 of 100 agreed, 1.000 against 0.90, made 2026-10-21

  9. The committee chair who stands as the supervisor, in the records of Digital Commons@Georgia Southern, Digital Commons @ West Chester University and University of Memphis Digital Commons, compared with the source's own pages by a program

    passes: 445 of 445 agreed, 1.000 against 0.90, made 2026-10-02

  10. The committee chair who stands as the supervisor, in the records of LSU Digital Commons, compared with the source's own pages by a program

    passes: 32 of 32 agreed, 1.000 against 0.90, made 2026-10-02

  11. The addresses formed for thesis pages, in the records of LUP, Lund, compared with the source's own pages by a program

    passes: 100 of 100 agreed, 1.000 against 0.90, made 2026-10-14

  12. What a format's reader reads, in the records of AMUR Poznan repository, compared with the records as the source served them by a language model, no person reading the sample

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-09-30

  13. What a format's reader reads, in the records of KU Leuven Lirias, University of Amsterdam (Pure) and UCLouvain DIAL, compared with the records as the source served them by a language model, no person reading the sample

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-09-29

  14. What a format's reader reads, in the records of Illinois IDEALS, compared with the records as the source served them by a language model, no person reading the sample

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-09-30

  15. What a format's reader reads, in the records of ETH Research Collection, KITopen and Knowledge@UChicago, compared with the records as the source served them by a language model, no person reading the sample

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-09-30

  16. What a format's reader reads, in the records of Aaltodoc, compared with the records as the source served them by a language model, no person reading the sample

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-09-30

  17. What a format's reader reads, in the records of White Rose eTheses, Harvard DASH and ERA Edinburgh, compared with the records as the source served them by a language model, no person reading the sample

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-09-29

  18. What a format's reader reads, in the records of LMU edoc, mediaTUM, OPUS4 Wuerzburg and OPUS4 Potsdam (publishup), compared with the records as the source served them by a language model, no person reading the sample

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-09-30

  19. The supervisors read, in the records of Purdue e-Pubs, compared with the source's own pages by a program

    passes: 129 of 129 agreed, 1.000 against 0.90, made 2026-10-12

  20. The supervisors read, in the records of NUŠL, the Czech national repository of grey literature; and Zaguán, Zaragoza, compared with the source's own pages by a program

    passes: 399 of 399 agreed, 1.000 against 0.90, made 2026-10-10

  21. The supervisors read, in the records of NUŠL, the Czech national repository of grey literature, compared with the source's own pages by a program

    passes: 149 of 149 agreed, 1.000 against 0.90, made 2026-10-10

  22. The supervisors read, in the records of Biblioteca Digital de Teses e Dissertações da USP, compared with the source's own pages by a program

    passes: 148 of 148 agreed, 1.000 against 0.90, made 2026-10-10

  23. The supervisors read, in the records of BCPW, Warsaw University of Technology; and Pomeranian Digital Library, Gdańsk, compared with the source's own pages by a program

    passes: 317 of 317 agreed, 1.000 against 0.90, made 2026-10-10

  24. The supervisors read, in the records of Repositorio UNAM, TesiUNAM, compared with the source's own pages by a program

    passes: 168 of 168 agreed, 1.000 against 0.90, made 2026-10-10

  25. Lists of names split into the people they name, in the records of Scholars' Mine, compared with the source's own pages by a program

    passes: 392 of 392 agreed, 1.000 against 0.90, made 2026-10-13

  26. Lists of names split into the people they name, in the records of Washington University Open Scholarship, compared with the records themselves by Claude, a language model, in a person's place on the maintainer's word

    passes: 120 of 120 agreed, 1.000 against 0.90, made 2026-10-08

  27. The level of each thesis, in the records of DSpace at University of West Bohemia, compared with the records as the source served them by a person, by hand

    passes: 50 of 50 agreed, 1.000 against 0.90, made 2026-10-02

  28. The copies a repository keeps of one thesis, in the records of NUŠL, the Czech national repository of grey literature, compared with the records themselves by a program

    passes: 74,805 of 75,022 agreed, 0.997 against 0.90, made 2026-10-10

  29. The supervisors read, in the records of Repositori UJI, compared with the thesis's own text by a program

    passes: 112 of 113 agreed, 0.991 against 0.90, made 2026-10-20

  30. The supervisors read, in the records of Vilnius University ePublications; and KTU ePubl, Kaunas, compared with the thesis's own text by a program

    passes: 226 of 229 agreed, 0.987 against 0.90, made 2026-10-10

  31. The supervisors read, in the records of Repositorio Aberto da Universidade do Porto and RENATES, compared with the thesis's own text by a program

    passes: 151 of 153 agreed, 0.987 against 0.90, made 2026-10-15

  32. The supervisors read, in the records of Vilnius University ePublications, compared with the thesis's own text by a program

    passes: 141 of 143 agreed, 0.986 against 0.90, made 2026-10-20

  33. The order of family and given names, in the records of DSpace at University of West Bohemia, compared with the records as the source served them by a person, by hand

    passes: 145 of 147 agreed, 0.986 against 0.90, made 2026-10-02

  34. The supervisors read, in the records of CaltechTHESIS, compared with the source's own pages by a program

    passes: 125 of 127 agreed, 0.984 against 0.90, made 2026-10-20

  35. The supervisors read, in the records of Minnesota Digital Conservancy, compared with the records themselves by Claude, a language model, in a person's place on the maintainer's word

    passes: 118 of 120 agreed, 0.983 against 0.90, made 2026-10-08

  36. The supervisors read, in the records of University of Amsterdam (Pure), compared with the thesis's own text by a program

    passes: 216 of 220 agreed, 0.982 against 0.90, made 2026-10-19

  37. What a format's reader reads, in the records of ScholarWorks at UMass Boston, CCU Digital Commons and OhioLINK ETD Center, compared with the records as the source served them by a language model, no person reading the sample

    passes: 49 of 50 agreed, 0.980 against 0.90, made 2026-09-29

  38. What a format's reader reads, in the records of HAL/TEL OAI, compared with the records as the source served them by a language model, no person reading the sample

    passes: 49 of 50 agreed, 0.980 against 0.90, made 2026-09-30

  39. The addresses formed for thesis pages, in the records of Qucosa, the Saxon universities, compared with the source's own pages by a program

    passes: 98 of 100 agreed, 0.980 against 0.90, made 2026-10-15

  40. The supervisors read, in the records of DSpace TUL, Liberec, compared with the thesis's own text by a program

    passes: 133 of 136 agreed, 0.978 against 0.90, made 2026-10-15

  41. The supervisors read, in the records of OpenBU, compared with the thesis's own text by a program

    passes: 130 of 133 agreed, 0.977 against 0.90, made 2026-10-20

  42. The supervisors read, in the records of TESEO, compared with the records themselves by a program

    passes: 352 of 361 agreed, 0.975 against 0.90, made 2026-10-21

  43. The committee chair who stands as the supervisor, in the records of Digital Commons@Georgia Southern, Digital Commons @ West Chester University and University of Memphis Digital Commons, compared with the source's own pages by a program

    passes: 372 of 382 agreed, 0.974 against 0.90, made 2026-10-02

  44. The supervisors read, in the records of Scholars Commons at Laurier, compared with the source's own pages by a program

    passes: 137 of 141 agreed, 0.972 against 0.90, made 2026-10-10

  45. The addresses formed for thesis pages, in the records of ERA, Alberta, compared with the source's own pages by a program

    passes: 96 of 100 agreed, 0.960 against 0.90, made 2026-10-14

  46. The supervisors read, in the records of ERA Edinburgh, compared with the thesis's own text by a program

    passes: 109 of 115 agreed, 0.948 against 0.90, made 2026-10-20

  47. The supervisors read, in the records of NUŠL, the Czech national repository of grey literature, compared with the source's own pages by a program

    passes: 133 of 142 agreed, 0.937 against 0.90, made 2026-10-19

  48. The committee chair who stands as the supervisor, in the records of Michigan Deep Blue, compared with the advisors Wikidata states by a program

    passes: 585 of 635 agreed, 0.921 against 0.90, made 2026-10-19

  49. The supervisors read, in the records of OpenAIRE: Biblos-e Archivo, UAM, compared with the records themselves by a program

    passes: 189 of 207 agreed, 0.913 against 0.90, made 2026-10-20

  50. What a format's reader reads, in the records of MIT DSpace theses, AUrora / Auburn University ETD, TDX (Tesis Doctorals en Xarxa) and Repositorio Aberto da Universidade do Porto, compared with the records as the source served them by a language model, no person reading the sample

    passes: 45 of 50 agreed, 0.900 against 0.90, made 2026-09-29

  51. What a format's reader reads, in the records of Leiden University Scholarly Publications and Radboud Repository, compared with the records as the source served them by a language model, no person reading the sample

    fails: 42 of 50 agreed, 0.840 against 0.90, made 2026-09-30

Some checks in words

Portugal's register of theses, RENATES, names the supervisors of many theses whose record at the University of Porto names none. Before the register's names were lent to Porto's records, a sample of them was looked for on the supervisor lines of the theses' own PDFs. At the Technical University of Liberec, the name read from a title page's "Vedoucí práce" line was compared with the supervisor the record itself names, on theses whose record names one. At Mexico's national university, a contributor listed with no word for their role was read as the supervisor only after the repository's own pages were found to list such people under its heading for advisors.

The first readers of several formats were tried by a language model with the raw records open, reading what the reader made of each record; no person read those samples, and the board says so. One of them failed: the reader for the format two Dutch universities serve, where the model found identifiers written as the word "null" and a field holding several people. Both were mended in the reader afterwards, but no later check of it is recorded, so it stands on the board as it was measured.

The joins of people across sources are held to the same bar, and are told on the page on who is who. One of them, a way of finding a Wikidata person's author record among OpenAlex's, fell short of the bar in each of the forms tried. None was let in on its sample. Instead every pair the best of them proposed was read on its own, against the person's Wikidata entry, by language models and no person; only the pairs read as one person were joined, and those the readers were unsure of are shown as likely, not confirmed.

Why a failed check is shown

A failed check is as much a finding as a passed one. It says what the record declined to claim, or claimed only after a reading was mended, and it shows that the bar is a real one: a method that never failed its checks would give a reader no way to tell a strict check from a lenient one. So every check is listed here as it was measured, the failures among them, and none is taken down when it fails.

Not everything has a check yet. Some rules for a source's habits stand on the maintainer's word after reading the source's pages, and others wait for a sample to be drawn and read; several readers were checked once, when they were first written, and not again. The board holds what has been measured and no more.

Every number traces to a reading

No count on the site is typed by hand. Each figure, the theses on the landing, the supervisions of the coverage page, the scores on this one, is counted by the site's own code from the record, and the answer it comes in keeps which reading of which source it was counted from. A page shows the day the reading was made, not the reading's own name, which says nothing to a reader; a citation of a person's page names the day the record was read in the same way.

That has two consequences. A number can be checked, by asking the same answer again or by following it to the source it counts. And a number changes when the record does: the figures here are the newest the record holds, and a later visit may find them moved.