Evaluate AI model distillation defenses with three ledgers: capability leakage, legitimate-user friction, and false-positive harm.