HOW SURE IS FEEDER?

The students a university recruits are about 18, but the first test counted everyone who leans toward it, alumni and sports fans included. Amendment 2 asked Qloo the same question about people aged 24 and younger, scored the same way on the same 109 universities, and was written down before any of it was fetched.

• It beats the AI assistant: +0.13 in 2022 (range 0.07 to 0.18) and +0.14 in 2020. What young people actually like explains more of where freshmen come from than what an assistant remembers.
• It beats the all-ages pull by +0.09 (range 0.05 to 0.12), in both years. The rule written in advance said the board would switch to it if so, and it has.
• A caution: another university's young pull also explains some (+0.14), so part of the gain is where young people are in general. A university's own young pull still beats that placebo by +0.26 (range 0.20 to 0.33).
• For ranking every state outright, the assistant is still better (0.85 against 0.78 with young pull), and young pull adds nothing on top of it (-0.00).

Ground truth: IPEDS, the federal record of where each university's first-time freshmen lived, by state. 124 universities that Qloo knows by name, with at least 1,000 freshmen and a fifth from out of state; ten universities tried by hand first were left out. The question is what Qloo explains beyond the obvious: big states and nearby states send more students.

Partial Spearman with what population and distance leave unexplained, averaged over universities, with 95% bootstrap intervals.

• Qloo's signal is real. It explains part of what geography misses (+0.32) in both years, and a placebo explains nothing.
• It matches the assistant there (+0.04 in 2022, range -0.01 to 0.09; +0.05 in 2020), from measured taste rather than memory.
• For ranking every state, the assistant is better, and adding Qloo to the assistant adds nothing (-0.00). Feeder's state board shows pull next to real enrollment so a team can see both; it does not claim Qloo beats an assistant at the ranking.

• Qloo's data is from 2026 and the enrollment data from 2020 and 2022, so this measures alignment, not a forecast.
• Where students come from also reflects past recruiting, alumni, athletics and price; the test can't separate taste from those.
• Metros, high schools, rival colleges and what resonates are untested. IPEDS has no detail below the state.
• Only first-time freshmen are counted, not transfers or graduate students.
• Qloo's age groups are its own estimates of who engages with what, and “24 and younger” is wider than freshmen.

This repository first held Shelf Desk, which picked teen books for students' interests from Qloo's book graph. Its pre-registered test on Seattle Public Library checkouts failed: per copy, its picks were borrowed 42.4% less than an assistant's in 2025 and 46.7% less in 2026, and no better than random picks from the same pool. It was dropped; the record is in docs/retired. Feeder exists because the test said to stop.