Every prediction, graded.

What is being recorded right now.

Every time the engine places a student against a school, it records a position: where that profile sits against that school's published admitted-student band, and whether the school reads as a reach, a match, or a safety. That placement is stored with the inputs and the published figures it was computed from.

The comparison set is the 394 US colleges that publish both a 25th and 75th percentile SAT range for admitted students. That is the corpus every placement is made against — it is a count of schools with usable published data, not a count of students or of scores produced.

Because the placement is arithmetic on published figures rather than a model output, it is reproducible: the same inputs and the same band return the same answer, and the record shows which band and which year it used. A prediction that cannot be reconstructed cannot be graded later.

How the grading works.

Admission decisions arrive in the spring. Students log what actually happened — admitted, rejected, waitlisted — for each school on their list, and that outcome is stored against the placement recorded months earlier.

In December 2026 we compare the two and publish the comparison. The unit is the individual placement: for schools we called a safety, what fraction were admitted; for schools we called a reach, the same. The misses are the point of the exercise, so the misses get reported at the same resolution as everything else.

We will also publish how many outcomes the report rests on. If the sample is small, that will be stated at the top rather than buried, because a rate computed on a handful of decisions is not yet evidence and saying so is part of the job.

Why industry claims cannot be checked.

Admissions marketing runs on multiples: some number of times better than the published acceptance rate, some percentage of clients placed at a top choice. Read closely, those figures are usually unfalsifiable — not necessarily false, but impossible to check.

• The denominator is chosen after the fact. A rate is only meaningful if you know who was counted, and clients who were screened out, who left, or who did not apply can be excluded without the number ever looking wrong.
• Selection does the work. A service that admits strong applicants will post strong outcomes with no effect of its own. Comparing its clients to the national acceptance rate compares two different populations.
• There is no prediction on the record. Outcome claims are made after decisions are known. Nothing was committed to in advance, so there is nothing that could have turned out to be wrong.

Our claim is built to fail if it is false. The prediction is recorded before the decision exists, the population is every placement rather than a chosen subset, and the report goes out on a fixed date whether or not it flatters us. That does not make it more accurate than anyone else. It makes it checkable, which is a different and smaller thing that we can actually promise.

What this will not become.

A graded record of positions is not an admission prediction, and the report will not turn into one. We will not publish a percentage chance of getting in, before or after December 2026, because published data cannot support it: essays, recommendations, institutional priorities, and everything else that moves real decisions appear in no public dataset.

What gets graded is the narrower claim we actually make — that your position against published admitted-student data is reported correctly, and that reach, match, and safety mean something when we say them.