The one piece of math worth understanding: bounded drift

The paper proves a result that is intuitive once you strip away the notation. Without governance, semantic drift accumulates with every update step, so after T steps the total drift grows in proportion to T. The longer the agent runs, the more corrupted its memory becomes, without limit.

Under SSGM, reconciliation runs every N steps and corrects the drift each time. So drift can only build up within a single window of N steps before being cleaned. The total drift is therefore bounded by the window size N, not by the full horizon T. The practical meaning: a governed agent's memory stays stable even over very long runs, while an ungoverned agent's memory degrades roughly linearly with time. That is the formal justification for why periodic reconciliation matters.