Submit your mixture for measurement
MoE model, mixture-of-models team, swarm, clan, council — if a group of models produces verdicts together, SwarmBench will measure what the group actually adds. Including when the answer is less than one.
What we need from you
Every member, with weights identity (hash or verifiable artifact reference). We join on weights, not names — a renamed model is a new model.
How a verdict is produced: router weights, voting rule, aggregation function. This is what delegation accuracy and consensus robustness are measured against.
You agree to the frozen-split law: practice items are visible, held-out items are never exported, the anti-Goodhart salt re-splits between generations.
The measurement contract
- We publish with CI and harness. Results ship with 95% confidence intervals, coverage fields (measured / unparseable / UNMEASURED), the exact harness version, and the corpus anchor hash. No number without provenance.
- Anyone can check it, free. Every published measurement carries its corpus anchor hash and harness version. Checking a published result costs nothing and needs no account.
- Negative results publish too. If your council measures n_eff below its seat count, that is the finding. It joins the Illusion Index next to ours. You can withdraw before publication; you cannot edit after.
No prices, no credits
There is no checkout, plan, tier or credit on SwarmBench. Verification of published results is free, forever, and a grade is never sold. A requested measurement never buys its result: the finding is published as measured, including when the answer is less than one.
Start
Commission a measurement by email — mention SwarmBench, your member count, and your weights references:
Agents: the machine-readable submission contract is in /agent.json and /llms.txt.