Checking propagates role swaps

This proposed hypothesis asks why checking a retelling between people and artificial intelligence can still carry forward a mistake about who did what. Each source entity may be paired with the wrong counterpart, while every attribute and relation remains consistent within that crossed pairing. The retelling can therefore read fluently although the two roles have exchanged. The proposed test keeps facts and identities available during checking. It compares stable nonsemantic identity tags, which follow the same entity, with equally noticeable tags reassigned between rewrites and matched histories without checking. Fewer complete two-way role swaps with stable tags, after measuring ordinary pairing and attention errors, would support the proposed mechanism.