About
How this archive is made, and how accurate it is
Software reads these pages out of old scans. It makes mistakes, we measure how often, and this page gives the numbers.
The short version
Every name, portrait and printed detail on this site was read from a scan of a book published in 1945 or earlier, by computer programs that read images and text (large language models and ordinary text-recognition software). People then review the cases the programs flag, and we audit samples to count the errors that get through. When we sampled newly added details against the books themselves, 96 of every 100 were true to the book (77 cleanly, 19 with awkward formatting) and about 4 were wrong. Home addresses checked against the page images were wrong in 1 of 120 cases. Directory households were wrong or garbled in 3 of 20. Nobody has checked every page.
Each person page names the book, the page number and a link to the scan, so you can check any fact yourself. If you find a mistake, tell us and we correct it against the scan.
Where the pages come from
All sources are open. The yearbooks, war rolls and biographical volumes are scanned by public and university libraries and held by the Internet Archive, the Library of Congress, Digital Commonwealth and similar archives. The city directories (Cambridge, Newton, Brookline, Quincy, Springfield and others) are digitised by public libraries and the Internet Archive. We do not use Ancestry, Fold3, Find a Grave, Newspapers.com or Classmates.com, and we do not copy from any paid or licensed collection.
The names in these books exist only inside the page images, so a search engine cannot find them. Reading them out is the work this site does.
Works published in 1930 or earlier are in the public domain in the United States. For books published from 1931 to 1945 we say only "Published YYYY"; we have not checked their copyright renewal, and our takedown process applies to them as to everything else.
How a person gets onto a page
- The scan is read. The text-recognition layer of the scan, the page image and a vision model are used together. The model proposes a name, the printed details, and where the portrait is on the page. A name is accepted only if it also appears in the page's own recognised text.
- The portrait is cut out of the page along the photograph's edges.
- Checks run on every portrait. A second, independent model confirms that the crop shows one person, with the head in frame and nothing else in it, and that the face is consistent with the first name (a man's photograph under a woman's name is the classic sign that a whole row of portraits has slipped by one place).
- Whole books are tested. For each book we measure how often face and first name disagree. A book above the limit is held back entirely, because a misread layout produces wrong faces in a pattern that single-page checks miss.
- Doubtful names go to a review queue and wait for a person. Two entries that claim the same portrait are both held. The same name with clearly different faces is held. Records with no source page are not published.
- Extra details (birth date, course, clubs, plans, prior school) are taken from the book's own text for that person. The first pass produced some errors, so a cleaning step removes the patterns we have found (ages mistaken for birth dates, roster headings, runs of OCR noise) and we re-audit.
- City directories are matched on the graduate's street address and surname. A page then says what the directory lists at that address. It never says "father" or "mother", because the directory does not.
- Notable people are linked to Wikipedia and Wikidata only when a second fact agrees with the book (for example the exact birth date), and the page shows the evidence.
A large share of what the programs read has been held back by these checks rather than published.
How accurate it is
These are the audits we have run so far. They measure specific things, so please read the scope column.
| What was checked | How | Result |
|---|---|---|
| Details added to existing pages (course, birth date, clubs, plans, prior school, address) | 100 random records from 100 books, 170 details, each compared with the book's own text for that person. 19 October 2026. | 131 correct as printed (77%); 32 true to the book but awkward, such as a hyphen left in a word or a role split from its club (19%); 7 wrong (4%). The margin is about 6 points either way. |
| By kind of detail, from the same sample | Wrong out of checked | Course 0 of 29. Birth date 0 of 28. Home address 0 of 16. Prior school 1 of 18. Plans after graduation 2 of 21. Clubs and activities 4 of 58 wrong, 19 more awkward. |
| Home addresses against the page images | 120 records from 20 books across Massachusetts; a person read each page image and compared the digits. 19 October 2026. | Digits wrong in 0. One address (1 of 120) belonged to the neighbouring person on the page. Formatting flaws such as trailing text or a split word in 15 (12.5%). With a sample this size, the true wrong rate could be as high as 4.6%. |
| City-directory households | 20 newly matched pages, each directory line read against the entry we show. | 17 clean; 3 of 20 showed a wrong or garbled entry, usually two directory lines merged. Relationships are never stated. |
| Portrait and first name agree | Measured for every book; the method with the highest rate is withheld. | Books with more than 1.5% disagreement (or fewer than 8 confident checks) are held back. |
What we have not measured: how often a surname is misspelled (the OCR sometimes turns "Vilette" into "Vllette" or cuts "Thomas" to "Thoma"), and how many faces are paired with the wrong name inside a book that passed. The audit numbers above apply to the details added to the first pages we built, not to every field on every page.
Mistakes we know about
- Spelling. Names, schools and mottoes can contain recognition errors; stray quotation marks and hyphenated line breaks sometimes survive.
- Labels. A grammar school can appear as a "course", an ambition ("To be a success in life") as a destination.
- Neighbours. On crowded pages a detail can come from the entry next to the right one.
- Directory households. Fathers and sons with the same name can be merged. A household lists the people the directory printed at that address in that year, not necessarily the graduate's family.
- Silent omissions. A detail the book printed may be missing from the page if it failed our checks. We would rather leave a field out than print one we doubt.
We correct what we find, and the list above changes as audits find new patterns. How corrections work is on the standards page.
What the programs are
The page-reading and checking steps use large language models from Google (Gemini) and Anthropic (Claude), plus the text-recognition layer that the Internet Archive already provides with each scan. No model writes any text that is displayed as the source's words. Where a page has a short summary of a notable person, it is labelled as a summary of the Wikipedia article and carries its licence.
Who runs it
The Class Album is an independent project run by one person. It has no institutional backing and is paid for out of pocket and by small affiliate commissions that never change what a page says (how the site is funded). Contact: hello@theclassalbum.com.
Check it yourself
Every page links to the original scan. If a name, face or detail is wrong, use "Something wrong on this page?" at the bottom of the page, or write to us. You can also help check names from the review queue; answers are never published until a second person agrees and we have checked.
Last updated: 20 October 2026. Audit dates and sample sizes are given above so you can see how old each number is.