What was wrong
When the site looks for a college's newest admissions figures, it finds the college's own documents and reads them. Some it found were never read:
- Common Data Sets with no year on the cover. Marquette's 2025–26 Common Data Set prints "2025-2026" on its cover but the words "Common Data Set" are a picture, so the year wasn't recognized and the file was set aside. About 50 colleges' files were in the same spot, most with the year only in the file name.
- Class profiles posted as PDFs were read as if they were web pages, so nothing useful came out of them.
- One odd link on a college's site (Western Carolina's home page has a template placeholder for a link) stopped the search for that college's documents entirely. 38 colleges were affected.
- A mistyped link. For Santa Clara, the search returned the right page and a link to its 2025–26 Common Data Set with one folder name wrong, so the file was never downloaded.
What changed
- The year now comes from the cover even when only the years are printed, or from the file name
(
cds-2025-2026_final.pdf). A file that states its year nowhere is read once under the current year, and its figures are published only if the file's own text confirms that year; otherwise they wait for a person to check. - PDF class profiles are read page by page, like other PDFs, and ones read the old way are read again.
- Links the search suggests are checked before they count: a page that doesn't exist is dropped and the search keeps going, and a mistyped Common Data Set link is replaced by the same year's link on the college's own list of files.
- An unreadable link is read as written instead of stopping the search.
These colleges are picked up by the next run.
Behind the scenes
Georgia Tech's published fall 2025 figures came from a class-profile PDF read the old way; the numbers aren't in that
file's text, so they should be checked by hand. The fixes are described in
the round 3 spec, and scripts/college-reported-rerun-ids.mts lists the
colleges to run again by reason.