The Sorting Machine
A data dossier

The Sorting Machine

What did SFFA reveal about “merit” in college admissions?

A domestic student at a given academic percentile has a far smaller chance of landing a seat at an elite university, and of paying for it, than an identical student fifty years ago. Admissions are fundamentally zero-sum, so where did these seats go?

One applicant enters at the top. The pegs are the factors: SAT, GPA, sports, legacy, donor, race. The machine sorts each drop into one of two bins, a wide REJECTED and a narrow ACCEPTED. Most land left.
01 · The squeeze

Different doors, wildly different odds.

Every applicant is sorted into a lane before the file is even read, and the gates on those lanes are not the same size. A recruited athlete walks through an opening five-sixths of the way open. The unhooked domestic applicant faces a pinhole. Same university, same year, a roughly 17-fold gap in the odds of getting through.

Admit rate by track: the gate you're sorted into
Applicants stream in from the left; each gate's opening is scaled to that track's admit rate. Watch where the crowd piles up.
SOURCE · Admit rates from Arcidiacono, Kinsler & Ransom, Legacy and Athlete Preferences at Harvard (using Harvard's records in SFFA v. Harvard): recruited athletes 86%, faculty/staff children 47%, dean's-interest 42%, legacies 34%, vs. under 5.5% for non-ALDC applicants. NBER w26316 · full paper. Lane widths (applicant volume) illustrative.
02 · The race gap

The same file, four different verdicts.

The most-cited number from the Harvard trial comes from a counterfactual. The plaintiffs' economist modeled one applicant profile and swapped only the race field. The model's predicted probability of admission moved from 25% to 95%.

Modeled admit probability for one identical applicant, by race
A male, non-disadvantaged applicant at a fixed academic and extracurricular profile, with only the race field varied.
SOURCE · P. Arcidiacono expert report, SFFA v. Harvard: a baseline applicant with a 25% chance as Asian American would have ~36% as white, ~77% as Hispanic, ~95% as African American. See also NBER w27068. Figures are the plaintiffs' model estimates, disputed by Harvard's experts.

This is a modeled probability rather than a real admit rate, and Harvard's experts disputed the specification. But no party disputed the direction: at equal academic strength, the personal-rating and race adjustments did not fall evenly across groups.

03 · Price discrimination

A different price for every family.

The published cost of attendance is a fiction almost no one pays. The real price is computed per household from a federal-and-institutional aid formula. This is textbook first-degree price discrimination: the same product, the same seat, sold to each family at the maximum the formula thinks it can extract. Here is the machine that sets your number.

1 · Cost of attendance
$97,985/yr
Tuition Housing Fees
The full sticker: tuition, housing, meals, fees, books, travel, personal. The number the formula starts from.
2 · Aid, set by your SAI
FAFSA + CSS Profile
The FAFSA (and, at elite privates, the deeper CSS Profile) turns your family's income and assets into a Student Aid Index. Grant aid = cost of attendance − SAI. The SAI is your price floor.
=
3 · Your price
$0 – $98k
Two students sit in the same lecture. One paid near nothing, the other near $98k. The seat is identical; only the family was re-priced.
What each family actually pays at Princeton, by income
Average net price after grant aid, 2023–24. The 9× jump at the $110k line, then the climb to the full ~$98k, is the cliff.
SOURCE · Avg. net price by income (Princeton, 2023–24): <$30k $41 · $30–48k $352 · $48–75k $1,217 · $75–110k $4,478 · $110k+ $36,094 — NCES College Navigator. Full cost of attendance ($97,985) — Yale, 2026–27. "Full pay" bar illustrative.

Why the upper-middle class gets hit hardest. The formula's design does the damage: income above a modest protection allowance is assessed at steep marginal rates, and the CSS Profile that elite privates use counts what the federal FAFSA does not: home equity in your primary residence and the non-custodial parent's income. The 2024 FAFSA overhaul also removed the reduction for having two children in college at once. Net effect: a $175k household with a house and a 401(k) is judged "full pay," pays near-sticker from already-taxed income, and receives essentially nothing. Dept. of Ed (SAI) · CSS Profile.

04 · Out of sync

Demographic comparisons.

Treat this as an accounting question. Follow each group across three steps: its share of the country's 18-year-olds, its share of all U.S. college students, and its share of an elite private class. If admission were neutral the three bars would stay level. They don't. Some groups expand sharply at the elite tier while others contract: Asian students roughly quadruple, international admits go from zero to a sizable slice, and several domestic groups shrink at every step.

The country → all colleges → elite, by group
Each group's share at three stages: U.S. 18-year-olds, all U.S. college students, and a stylized elite private class.
SOURCE · College-age population (18–24, 2023): White 51.7% · Hispanic 24.2% · Black 13.7% · Asian 5.8% — NCES Digest tbl. 101.20. Elite shares from Harvard CDS 2023–24 (White 31.8% · Asian 23.7% · Intl 13.9% · Hispanic 11.7% · Black 9.5%) — Harvard CDS. Race×gender split within White is estimated.

The hook paradox.

Here's the counterintuitive part. A group that leans heavily on legacy and recruited athletics doesn't get those hooked seats for free. They come out of the group's overall footprint. So when you strip the hooks away and look only at the unhooked "merit" pool, that same group can be crowded out of it: its own hooked admits have already spent the seats. First, who actually fills each track:

Who fills each admissions track
Estimated racial composition of each track's admits. The hooked tracks and the unhooked pool have very different makeups.
SOURCE · Recruited athletes, legacies, and dean's-interest admits are each over 68% white; non-ALDC applicants are under 41% white. ~75% of white ALDC admits would have been rejected without the tip, and over 43% of white admits are ALDC — Arcidiacono, Kinsler & Ransom, NBER w26316; Harvard Crimson. White shares are sourced; splits among the non-white remainder are estimated.

The merit ratio makes the paradox one number per group: its share of unhooked admits ÷ its share of all admits. Below 1.0 means the group is thinner in the merit pool than on campus overall, because the hooks did the heavy lifting and its unhooked applicants face the steepest pure-merit climb. Above 1.0, the group earns its place and would gain seats if hooks were abolished. Split it by race and gender and the pattern sharpens: men lean harder on the male-skewed athletic and legacy hooks, so every man sits below his female counterpart on the same scale.

The merit ratio, by race and gender
Each group's share of the unhooked (merit) pool ÷ its share of all admits. 1.0 = neutral; left = hook-dependent, right = merit-carrying. Men trail women in every group.
SOURCE · Method as in the track composition above, with the male skew of athletic hooks applied (recruited athletes run roughly 55–60% male). The race-and-gender split is the most modeled figure on the page: a directional estimate, not a published statistic.

And per head of population? Divide each group's unhooked merit seats by its share of the country's 18-year-olds, so that 1.0 now means merit seats in exact proportion to that group's youth. The merit pipeline turns out to be a rounding error for most Americans and a firehose for one group:

Merit seats per 18-year-old, by race
Unhooked (merit) seats going to each group ÷ that group's share of U.S. 18-year-olds. 1.0 = population parity.
SOURCE · Denominator — U.S. 18-year-olds by race (White 51.7% · Hispanic 24.2% · Black 13.7% · Asian 5.8%), NCES tbl. 101.20. Numerator — each group's share of the non-ALDC (merit) admit pool, built from Harvard CDS demographics and the ALDC racial skew (hooked tracks >68% white). A derived index, not a single published figure; international excluded (no domestic base).

The hook advantage and the merit disadvantage are the same coin. The more a group is admitted through legacy and athletics, the more of its overall footprint is already spent, so its unhooked applicants compete for a shrunken remainder of merit seats, and against the largest unhooked applicant pool. The group that looks most "privileged" by the hooks is the one whose merit pipeline is squeezed hardest. (This reads the trial data as schools managing each group's total toward a rough band, which is well-evidenced but a point critics contest.)

05 · The vanishing test

Two quiet thumbs on the scale.

Since 2020, two changes have narrowed the merit door further. Both were sold as fairness. Both systematically disadvantage the same person: the high-scoring, unhooked applicant who has nothing but a strong record to stand on.

Test-optional: more applicants, the same seats
Harvard undergraduate applications, in thousands. Test-optional began for the Class of 2025, and applications jumped ~50% while the class stayed the same size.
SOURCE · Harvard applications: 40,248 (Class of 2024, last test-required cycle) → 61,220 (Class of 2026, record-low 3.19% admit rate) — Harvard Crimson. Share of Common App colleges requiring scores fell ~55% (2019–20) → ~5% (2024–25); >1,900 colleges now test-optional — FairTest.

Why "optional" hurts the high scorer. A 1550 used to be a cheap, legible signal that vaulted an unknown applicant out of the pile. Make the test optional and that signal goes quiet: a top score no longer separates you, because the reader leans harder on essays, activities, and "context," the soft factors that reward the polished and the coached. A Dartmouth study found disadvantaged applicants scoring above 1400 were 3× more likely to be admitted when they submitted scores. The very students test-optional was meant to help are hurt when they hold a strong score back. NBER Digest, 2025.

The big-fish penalty.

Every file is read "in context" of its high school. Reasonable in theory, but it means a strong student at a strong school is measured against a wall of equally strong peers, while an identical record at a weaker school stands out. Percent-plans and holistic "context" reward being a big fish in a small pond, and quietly penalize the applicant who swam in a harder one.

Same 1500 SAT, four different high schools
Estimated admit odds for one identical applicant, varying only the competitiveness of the high school.
SOURCE · Illustrative model of the school-context effect, following percent-plan and holistic-review research (Texas Top 10%, UC ELC). Directional, not a measured admit rate.

Stack the two together and they point at one person. The applicant with the strongest raw record, a top score earned at a rigorous school, is exactly the one the test-optional shift de-emphasizes and the context lens penalizes. Precisely the student the word "merit" was supposed to name.

06 · The counterfactual

The same student, three admissions systems.

Here is a hypothetical applicant, call him Daniel R., a strong, unhooked domestic student. He is not a real person, but a composite built to hold academic strength constant while we change only the system he applies through. What determines his odds is less his file than the machine reading it.

SAT 1530 (~99th pct) GPA 3.95 UW Intended major: Engineering No legacy · no recruited sport · full-pay-adjacent
United States

Holistic review

One reader scores the whole person: essays, personal rating, "institutional priorities." Academics are necessary, not sufficient.
Est. admit odds, top-tier
~5%
A near-perfect academic file only gets you to the baseline. Unhooked, male, and full-pay competes hardest against itself.
United Kingdom

Course-based selection

Applies to one subject at Oxbridge. Judged on predicted A-levels, a subject admissions test (e.g. the ESAT/MAT), and a technical interview.
Est. admit odds, at course level
~20%
Demonstrated subject ability dominates. Personal "fit" and extracurricular breadth carry little weight.
France

Concours (exam)

Two years of classes préparatoires, then a national competitive exam. A rank on a written test allocates the seat.
Determinant of the seat
Rank
Almost purely exam-meritocratic at the point of entry. No essay, no personal rating, no legacy channel.
Build a student · run the machine
Estimated admit odds
36%
Estimated net price / year
$84k
ILLUSTRATIVE MODEL — a simplified toy calibrated to the public figures on this page (Arcidiacono race deltas; net-price-by-income bands). Not a prediction for any real applicant or school.

Every one of these systems carries its own biases upstream. The deeper point is that "merit" is defined by the machine that reads the file. Move Daniel across the border and the very traits that make him a residual in one system make him a front-runner in another.

07 · The punchline

Same student. Fifty years apart.

Stack the four forces together, the reserved seats, the holistic filter, the per-family price, and the demographic targets, and they compound. Hold the student constant at a top-decile academic profile and ask two questions across half a century: can he get in, and can his family pay for it? On both, the line moved the wrong way.

The odds on a top-decile domestic applicant, then vs. now
Two measures for the same academic percentile: chance of admission, and net price as a multiple of median household income.
SOURCE · Admit rates from historical institutional reporting vs. current Common Data Sets; affordability indexed to Census median household income. A directional composite meant to show the trend, not exact endpoints.

Fifty years ago a strong domestic student faced a selective but winnable gate, at a price a middle-class family could absorb. Today the same student faces a lottery, and if he wins it, a bill indexed to a wealth he doesn't have. The pipeline didn't close. It narrowed to a trickle, and got expensive.

Methodology & sources

A note on the numbers.

Editorial note: the counterfactual student Daniel R. is a composite constructed for illustration and does not depict any real individual.