How the Ivy Admit essay scorer was built — and what it can and can't tell you

The training corpus

The scorer is anchored on patterns observed across roughly 500+ accepted personal statements and supplemental essays gathered from students who shared their applications publicly (anthologies, university-published samples, alumni archives) along with structurally-comparable rejected drafts. We do not redistribute the underlying essays. The model uses them only as alignment signal during evaluation — it reads patterns, not paragraphs.

Schools represented in the corpus include Harvard, Yale, Princeton, Stanford, MIT, Columbia, UPenn, Brown, Dartmouth, Cornell, Duke, Northwestern, UChicago, Caltech, Johns Hopkins, Rice, Vanderbilt, Notre Dame, Georgetown, Williams, Amherst, Pomona, Swarthmore, Bowdoin, the UCs, and the Ivies' peer set. The corpus skews toward humanities and social-science topics; STEM-leaning essays are validated separately against a smaller sample to catch evaluation drift.

The rubric

Every essay receives three orthogonal scores on a 0–100 scale:

• Content (specificity).Density of concrete, nameable detail relative to the essay's claims. Penalizes claim-without-scene, abstract aspirations, and sentences that could appear unchanged in another applicant's essay.
• Structure (narrative arc).Whether the essay moves: opening scene → moment of change → forward-looking close. Penalizes summary openings, midstream stall (no inflection point), and resolved-growth closes (“this taught me…”).
• Voice (distinctiveness).Lexical and syntactic distance from the modal college essay. Penalizes high-frequency template phrases, adjective stacks, cliché openings (quotes, “ever since I was young”), and overly uniform sentence rhythm.

Composite scores are weighted by essay length: longer essays weight Content more heavily; shorter supplements weight Voice. Sub-scores never feed each other — Content failure cannot be compensated by Voice strength.

Validation

We validated rubric scores against admit-outcome buckets in two ways:

1. Held-out admitted-vs-rejected pairs. On a held-out test set of admitted and rejected essays at peer schools (matched for academic profile), composite scores separated admitted essays from rejected ones at a meaningful but non-determinative rate. Essays from admitted students at highly selective schools average above 82 across the three dimensions; the spread is wide and includes admitted essays scoring in the 60s.
2. Inter-rater agreement.A panel of former admissions readers and college counselors independently rated a sample of essays on the same rubric. Composite scores correlated with the panel's ranked order at roughly the same level that two human readers agree with each other on a single essay.

We re-run validation when the model is updated. Recent results are available on request to support@getivyadmit.com.

Known limitations

• The essay is one signal.Admit decisions also weight grades, course rigor, recommendations, activities, demographic context, and institutional priorities the application can't see. A high score does not predict an admit; a low score does not predict a rejection.
• Topic-area drift. The corpus skews humanities. STEM essays (research-heavy, technical) are scored against a smaller calibration set and may under-weight a particular kind of intellectual specificity.
• Non-native-English drafts.Voice scoring is calibrated to native English prose patterns. We are working on a separate calibration for non-native writers; in the meantime, the Voice score is less reliable for essays where English is the writer's second or third language.
• School-specific fit.The scorer evaluates an essay against general accepted-application patterns, not against any one school's admit profile. A strong essay can still miss on school fit, prompt-specific strategy, or supplement sequencing.
• Hooks and institutional priorities. Recruited athletes, legacies, first-generation applicants, and applicants advancing institutional priorities are evaluated in a different pool. Headline acceptance rates compress these signals; published-rate comparisons (used in the odds calculator) are mathematically less meaningful for hooked applicants.

Data sources for the admit-odds calculator

The per-school admit-odds calculator uses publicly published data from:

• U.S. Department of Education College Scorecard for acceptance rates, test-score ranges, financial aid, demographics, completion, and earnings.
• IPEDS (Integrated Postsecondary Education Data System) for the underlying federal data — enrollment, retention, graduation, and outcomes.
• Each school's most recent published Common Data Set (2024–25) for cycle-specific admit-cycle stats and demographic breakdowns.
• School-published news releases for current-cycle acceptance rates where Common Data Set data is not yet released.

Data is verified against source on a monthly review cycle. The footer of every comparison and college page lists the last verification date.

Privacy

Essays you submit through the free scorer are processed in-memory for evaluation and are not used to train models. Saved essays in your account are stored encrypted at rest and are private to your account; you can delete them at any time. We do not sell or share user-submitted essays under any circumstance.

Updates and corrections

Corrections, methodology questions, and data-source disputes: support@getivyadmit.com. We publish material methodology changes in the changelog and update this page when the rubric, corpus, or data sources change.