Ampersand Prep

This project is a four-agent stack that rates SAT question difficulty the way students experience it. Frontier models solve nearly any SAT question but cannot reliably judge how hard it is for a seventeen-year-old. Difficulty labels decide what practice a student receives and, on the digital SAT, drive the adaptive second module. An “easier than it is” label puts students in front of material they cannot use and blames them for the failure.College Board sets final Easy/Medium/Hard ratings from real test-day data. Practitioners see only the finished labels. In practice, difficulty for practice material is a tutor’s or AI’s guess. This project builds a system that rates from the student side and tests it against College Board ratings, treated as ground truth.A formal baseline was run on the full College Board bank (3,770 questions: 1,845 R&W, 1,925 Math; 1,296 Easy / 1,254 Medium / 1,220 Hard). A fresh Muse agent (no project context, calibration, or memory; questions shuffled and IDs stripped) answered 96.6% correctly, identified domain 97.6% and skill 93.3% of the time, but matched difficulty on only 50.8% (κ=0.256). It recognized Easy 91.2%, Medium 38.6%, Hard 20.4%. Of 1,220 Hard questions it called 793 Medium and 178 Easy. Its distribution was 2,118 Easy / 1,390 Medium / 262 Hard (average signed error −0.47 bands). Solving and judging are different skills; the error runs toward easier.