Quad

New featurePull request #54

The newest admissions figures, everywhere, and the first 22 are in

The newest admissions figures a college has published are now the ones the site shows everywhere, with the source in each value's ⓘ instead of tags; 22 colleges have them already, and collecting them costs about a tenth of what the first run did.

What's new

  • The newest numbers, everywhere. When a college has published admissions figures for a newer class than the federal data covers, those are the figures the site shows: on its profile, in Explore and Compare, and in rankings and medians. Applicants, admitted, enrolled, the admit rate and yield each describe the newest class the college published them for, and each one's ⓘ says which document it came from, the exact sentence, the link, the year, and the federal figure it replaced.
  • A calmer page. The green source tags and the "Figures marked like this come from…" line are gone. The ⓘ next to a number holds everything about its source. On Compare, a value from a different year than the rest of its row shows the year in small text.
  • 22 colleges already have newer figures. Ten from their 2025–26 Common Data Sets (fall 2025 class): Baylor, Berkeley, Columbia, Harvard, Michigan, MIT, Texas Christian, USC, Vanderbilt and William & Mary. Twelve from class profiles: Duke, Illinois, Loyola Chicago, Spelman and Villanova for fall 2026; Georgetown, Loyola Maryland, Morehouse, Pepperdine, Purdue, Stanford and UT Austin for fall 2025. A class profile often gives only part of the funnel; the rest stays federal, with its own year in its ⓘ.
  • Years can differ between colleges shown side by side, since each publishes on its own schedule. The Data page explains the rule.

Behind the scenes

  • The first 50-college pilot cost about $80 (and hit the account's credit limit before finishing), because discovery read whole documents with web fetch, resent them every turn, and retried blocked sites with the most expensive model. Discovery now returns links only, with capped page reads, streaming, one retry, and a low effort setting; next year's Common Data Set URL is guessed from last year's before any model is called; a blocked site goes straight to the review queue as "unreachable" with no model call; a failed check gets one re-read by Sonnet 5, and Opus is no longer used. Expected cost: about $0.05–0.12 per newly discovered college, under $5 for a quiet monthly run.
  • A dollar cap per run (--max-cost, $25 by default for manual runs, $150 for scheduled ones) stops a run cleanly, keeping every college it finished. The pipeline also saves after every college, logs its running cost, stops at once when the account's spend limit is reached, and still opens a pull request after a stop or cancel.
  • The data pull request now carries the site's data. npm run merge-reported folds the collected figures into the dataset with the same code the sync uses, the workflow runs it, and the PR's preview shows the figures. npm run report-college-reported prints which colleges have what.
  • Two pilot misreads were corrected: a Common Data Set whose web address included an upload-date folder was read as the wrong year (Vanderbilt relabeled fall 2025; Cornell dropped, since its existing CDS already covers that class). The year now comes from the file name first.
  • Specs: specs/college-reported-round-2.md (what the pilot showed and the decisions), with college-reported-data.md, data-lineage.md, data-page.md, school-profile.md and college-reported-setup.md updated.