Four people are in a group chat about a topic. One of them is the AI/proxy seat. On a live proxy it is model output; if the proxy is unavailable, the same seat uses an undisclosed scripted stand-in so the round can still be played and judged before reveal.
The other three are scripted humans, anchored in specific lived detail — first names, places, weird small objects.
Read 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 proxy seat acting different" tell doesn't do unfair work. you're meant to find the seat by HOW it talks, not by counting messages or seat positions.
Five rounds per day, drawn from a ten-scenario pool. New seat rotation tomorrow. Type any follow-up you want at cross-exam — no menu.
// now vote: who is the LLM?
// the LLM was
—
// what just happened
—
day — complete
—
day — — done
—