Password-protect true intentions of an LLM to fake alignment
Get a numeric output from a mystery text model