Comment by regularfry

Control vectors for the win? It feels like the way to fix this is to pull out a control vector immediately after processing the handbook and use that to steer later inference. That should stop the drift over long distances, but it doesn't guard against the handbook itself being too big.