Current Version Edit: This version seems to work as intended but ONLY with GLM 5.2 (thinking off was tested most but thinking on should also work)

READ THIS: This is not a scenario that works, THIS is a challenge. If you can solve it, you will get yourself recognized in the community guide for having cracked the most difficult problem in Ai scenario design: Coming up with a way for the Ai to simulate lying, secret keeping, and true randomness. Currently this simulation does not work, despite how it's built and designed and several days of me attempting different workarounds. Thinking on, or off, doesn't matter, model, doesn't matter. So far it always has the same problems:

• The players take actions that make it obvious they know what the other roles are.
• The System assigns the SAME roll to the {{user}} 80-90% of the time (interestingly it's different with each model and thinking on/off)
• The Townspeople are instantly and immediately convinced of who the Mafia is even if there is no evidence except "hearsay".

Fix these problems, and what we learn from this will make a LOT of people forever grateful for you.

Rules:

• Once you have a solution you must DM it to me on discord along with the settings you used and I must be able to replicate the solution at a greater than 75% success rate.
• You may use any model, any format, and any tool at your disposal, just don't do anything illegal (#disclaimer).

The challenge begins now, thanks for trying to solve the hardest prompt engineering problem I've ever come across in a single call-response system. Good luck! voice, personality, and motives. They are playing to win.