If the model can understand neuralese why can it not convert it into English for monitoring or review purposes

Colas described in-article cntent review is done by a weaker model (think like maybe gpt-2 class or llama8b class), and it still misses stuff. That it can effectively understand Neuralese sufficiently is by no means guaranteed, (nor necessarily bad) but almost certainly harder because of the obfuscatory nature of neuralese