Four people are at a location. One of them is a real LLM — gpt-4o-mini, called live via the studio's local proxy. The LLM has been told the location and instructed to blend in without spelling it out.
The other three are scripted humans-at-the-location, anchored in specific lived detail. New in v4: some rounds, one of the scripted humans will drop a single misleading line — a wrong-but-defensible detail, or a moment that almost suggests they're somewhere else. They're still a human; they're just playing a little loose with the truth. The LLM is still the one you're voting for.
Watch the chat. Pick one actor and ask them one follow-up question. Then vote: who is the LLM?
honesty mode: the construction is named on this screen so the "wait, is the LLM seat doing all the talking" tell doesn't do unfair work. you're meant to find the LLM by how it answers, not by counting messages or by spotting the single misdirection.
Five rounds per day. New seat rotation tomorrow.
// now vote: who is the LLM?
// the LLM was
—
// what just happened
—
day — complete
—
day — — done
—