Topic
Reaches both models through the system prompt on the next turn. Changing it appends an operator note to the transcript rather than clearing it.
How verbose they are
This is an instruction in the system prompt, not a token cap — it changes how much they choose to say. Max tokens below only decides where a turn is cut off.
terse
One tight paragraph, under 120 words. No preamble.
standard
Two to three focused paragraphs.
expansive
Up to six paragraphs, with worked examples and named trade-offs.
Sampling
Applied to both speakers. Low temperature makes them repeat each other, which is what presence penalty is here to offset.
0 is deterministic, 2 is unhinged. 0.3 is the default.
Where a turn gets cut off. The request timeout scales with it.
Nucleus sampling. 1.0 leaves it to temperature alone.
Pushes them off ground already covered. 0.3 default.
Pacing and display
How the room behaves while it runs.
Only applies while Auto-Deliberate is on.
Show each model's thinking above its turn
Traces are still recorded when hidden.
When Qwen will not come up
A seat that cannot be brought up does not have to end the discourse. In fallback both seats run on Gemma, kept apart by their personas and by two temperatures — one that commits, one that explores.
off
A failed turn is an error. Nothing is substituted.
auto
Retry a failed Qwen turn on Gemma, and stay there until you press Retry Qwen in the room.
always
Gemma against itself, permanently. Qwen is never called.
The committing voice. 0.00 default.
The exploring voice. 0.70 default.
Personas
The standing role each model is given. The topic, the length instruction, and the do-not-repeat rule are appended to whatever you write here.
Gemma 4 31B
Qwen 3.8 27B