Mixture IRT: The Safety Valve for Guessing

Interactive companion to XXX (2026)

All plots show performance relative to Oracle — zero is optimal (black line), more negative is worse.

Study 1 · DGP: Mixture · Fitted: Rasch, 1PLg, Mixture, 3PL

0.05 0.5
0.5 0.99
500 5000
10 60
20 80

Orange (1PLg): Applies fixed floor. Crashes away from blue line.

Green (Mixture): Tracks Oracle regardless of g_true.

Purple (3PL): Item-level guessing — modest, variable gain.

Grey (Rasch): Stable but suboptimal baseline.