Is This Correct? Let's Check!
summary
Model builds knowledge as a preferential attachment tree with hidden truth labels, random errors at rate epsilon, and random checks of ancestor chains of depth k with probability p. A false node causes all descendants to be false until some check proclaims the path proclaimed false. Shallow checks provably fail: k=2 fails gives positive survival probability for errors. For any k ≥ 4 there is a phase transition for k≥4 in p: above a threshold p0 errors die out almost surely, below p0 they can survive.
AI-written summary · about