SwarmBench — measured collective behaviour, not coordination vibes

Council of Sovereign AI (CSOAI) · protocol v0.1 · status: harness + protocol published, measured results land here

The gap this fills

The 2026 multi-agent benchmark field (MARBLE, CollabBench, HAS-Bench and peers) measures coordination quality. Nobody publishes, as one recomputable artefact:

1. Collective uplift — swarm score minus best-single-agent score on the same frozen split. If uplift ≈ 0, the swarm is one model with extra steps, and the card must say so.
2. Effective sample size (n_eff) of the swarm's decisions — correlated members voting is not N independent opinions. n_eff is reported next to every accuracy number.
3. Bootstrap 95% CIs on every metric, item-resampled, seed-pinned.
4. Fault injection — members are degraded/dropped in controlled patterns; the metric is degradation slope, not uptime.

The honesty clause (binding on all SwarmBench publications)

SwarmBench exists because of a retraction: CSOAI's own 33-agent council was retracted when measured n_eff came back 1.21 — thirty-three correlated agents voting was one opinion with extra latency. That retraction is part of this protocol. Any result published under SwarmBench must include the n_eff at which it was measured, and any swarm whose uplift CI includes zero must be reported as no measured collective benefit.

Harness contract

• Frozen item splits, published with the run — no adaptive item selection.
• Deterministic seed; reruns must reproduce within CI overlap.
• Every run emits a signed JSON artefact (Ed25519) so third parties verify the run happened as reported rather than trusting a table.
• Register: measurement and attestation vocabulary only. No "emergence", no "swarm intelligence" claims, no accuracy numbers without CIs.

Status and roadmap

• ✅ Protocol published (this card) · landing site: openmoe.ai
• ✅ n_eff measurement harness runs in CSOAI CI (see csoai/csoai-benchmarks n_eff_* artefacts)
• ⏳ Reference harness (swarmbench.py: uplift + bootstrap CI + fault injection, with the register-locked public sentence printer) ships here next, with the first measured swarm battery on open-weight clans
• ⏳ Public leaderboard opens only after three measured swarms with non-overlapping uplift CIs — a leaderboard over noise is worse than no leaderboard

Citation

If you measure swarms with this protocol, cite the artefact version you used and publish your frozen split. Measurement that cannot be recomputed is marketing.