Data, and where it breaks

Common Data Set parsing — measured, not estimated

Section C7 is a 19×4 grid in which a school states how much each admissions factor matters. It is a published, per-school, ordinal statement of the institution's own weighting function, and almost nobody uses it quantitatively. It is also a table in a PDF.

The dangerous failure is not an exception. It is silent column misalignment producing a plausible-looking row of numbers that is wrong. Two such bugs were found during development: a C1 parser that reported the year "2023" as an applicant count, and column detection in which "Very Important" contains the word "Important", quietly shifting every assignment on the page. Both now fail loudly instead.

Discovery, not parsing, is the bottleneck. Blind probing of the paths institutional research offices use returns almost nothing; the Internet Archive's CDX index — the only way to recover years a school has since withdrawn — was unreachable during this build.

UNC: the second school we cannot read

These tables are rendered as raster images in the filed PDF. pdftotext recovers the caption, the notes and the source line, but not a single cell. Confirmed by extraction, not assumed.

Consequence: UNC is NOT available as a second Tier-A shape source at launch. Until these tables are transcribed by hand (queued), Harvard is the sole individual-data-derived shape in this model, and the Harvard<->UNC cross-transport test in validate/transfer.py cannot be run in full. The site must not claim two schools of individual-level grounding.

Status: queued for manual transcription · ~5 tables x ~30 rows, roughly 2-3 hours of careful hand entry, or an OCR pass with two-of-three agreement checking

The staleness problem

The only coefficients estimated on real applicant records come from the Classes of 2014–2019 — application cycles ending around 2015. That is before test-optional admissions, before the Common App volume surge, and before the 2023 ruling. Harvard's admit rate then was about 7.4%; it is now about 3.5%. The level transports badly, which is why every school is re-anchored to its own published rate. The shape transports better, but how much better is exactly what remains untested.