All answer; the highest-weight member merges them into one.
Judge-best
All answer; the experiment's trusted judge picks the best.
Vote
Each generator scores the others on the rubric; the top weighted tally wins.
Consensus
Answers are clustered; the highest-weight cluster's answer wins.
Mixture of experts
A router picks ONE model to answer — the others are never called.
Match first; ask the judge only on a tie. Each prompt is compared by meaning with every member's expertise description from the model catalog. One clear winner answers straight away, at no extra cost. When two or more are too close to call, the experiment's trusted judge is asked — at temperature 0 — which of them suits this prompt best. A tie the judge cannot settle goes to the highest weight, never to chance.
One generation per round, however many models are in the configuration; the routed answer is scored exactly like a single model's. The match and tie thresholds are install settings (MOE_MIN_SIMILARITY, MOE_TIE_MARGIN), not per-configuration fields.
ModelTempWeight
claude-opus-4-8aggregatortop tier
anthropic
Long-form reasoning and careful multi-step analysis over large documents.
gpt-5
openai
Fast, broad general knowledge; strong at code and structured output.
All members are generators; the highest-weight one is the aggregator under Synthesize, and the strong tier plus the router’s tie-break under Mixture of experts. A single configuration keeps just one model; every other method needs at least two.