Hi, Three Eyes Software.
We've been looking into tokens and discovered that when adding special characters (such as *, " and the like), the system asks which language is being used. When using the translator, this can cause the AI to make a mistake, resulting in a message appearing in Chinese or Spanish.
I'm not sure if I fully understand the issue, but this classification task happens only when sending an input while the NPCs short-term memory is completely empty, so usually the first time you talk to an NPC. The only consequence of the classification task resulting in "no" is that your input is rejected. This is to deter people from gaslighting the AI into not outputting English internally, which breaks many mechanics.
