Game modes
Last edited by · ·
Pick a mode when you start a game. The AI players are always chosen separately.
| Mode | ID | For | Speaking order | Length |
|---|---|---|---|---|
| Play · LLM host | human-llm | 1–9 people, AI fills the rest | AI host picks speakers | up to ~43 min |
| Play · Classic host | human | 1–9 people, AI fills the rest | Simple rotation, no host model | up to ~43 min |
| Watch · 9 AI · Human-paced | standard | spectators | Host picks speakers | up to ~38 min |
| Watch · 9 AI · Fast-paced | fast-llm | spectators | Bids decide who speaks | up to ~36 min |
| Test · 9 AI · Fixed setup A2 | reproducible | developers | Host picks speakers | up to ~38 min |
| Test · Scripted verification | smoke | developers | Bids | minutes |
Every mode uses a random NewD3 setup except reproducible, which fixes setup A2 and the
seed. The model responses themselves still vary.
Timing
The play modes, standard and reproducible share human pacing:
| Phase | Time |
|---|---|
| Day discussion | 150 s (eleven 13 s speaking turns, then a 7 s turn) |
| Vote | 45 s |
| Night coordination | 30 s |
| Night actions | 45 s |
| Dusk / dawn | 5 s each |
Play modes also wait for people: the game starts when all nine seats are connected, or five minutes after the first person joins.
fast-llm uses fixed 10-second windows: 26 per day/night cycle (12 private daytime
coordination windows, 6 public bid rounds, 1 vote, 6 night coordination windows and 1 night
action). A full cycle takes 4 min 20 s. Windows never close early, even when everyone has
answered.
Every game ends early when a team wins, and is a draw after eight nights.
The host
In the paced modes a neutral host picks who speaks next. It never knows anyone's role.
- LLM host: a model picks the next speaker, prioritizing unanswered questions from people and fair airtime. If it fails or is late, a simple rotation takes over.
- Classic host: a predictable rotation that puts people's mentions first. No model calls.
Developers can override this with moderator in the game config: default (classic), llm
(require the LLM host) or auto (use an LLM when configured). The fast and smoke modes
always use bids and reject llm.