Quad

ImprovementPull request #58

College figures publish as soon as they pass their checks

Figures that pass their checks now publish on their own, a few checks learned from the first live runs, and discovery is tighter and faster.

What's new

  • Whatever passes, publishes. A run no longer waits for a person because some colleges couldn't be read. The figures that passed every check go live once the automated checks finish; the ones that didn't wait in the review queue. The only hold left is for a run in which many already-published figures suddenly change, which would point to a reading error rather than new data.
  • Rounded rates are accepted. A college that prints its admit rate as "42%" is no longer rejected because its counts work out to 42.43%.

Behind the scenes

  • The circuit breaker keeps only its change-share trigger (more than a quarter of already-published values changing in one run). Its failure-share trigger fired in both live runs on blocked sites and rounding and never on a real problem.
  • Auto-merge is on by default for manual runs as well as scheduled ones.
  • The rate check allows 0.5 points for a whole-percent rate, 0.1 for one decimal, 0.05 for two.
  • When discovery returns last year's Common Data Set link, next year's URLs are tried beside it.
  • The document reader's output limit doubled (a long PDF hit it); discovery uses fewer, smaller page reads and is cut off after six minutes per college; run summaries now break out cache reads and web searches.
  • Measured on the first ten-college run: 26.5 minutes and $1.93, with 4 of 10 colleges publishing. Details in specs/college-reported-round-2.md, "Round 2.1".