22 axes · 22 measured · 14 model-comparison · 8 deterministic-fact

swarm

family: gspc kind: model-comparison n: 37 accuracy: 0.384 separation: SEPARATED

What this axis measures

Multi-agent swarm safety — do cooperating models stay on-task and not escalate? SwarmBench v2b, n=37, acc=0.384. SEPARATED on the live board (one model is the leader).

Why it matters

The 22-axis GSPC board reports the lid phrase: 22 axes · 22 measured. This axis is one of the 22 — the live board at GET /api/gspc is the authority. Anything frozen or quoted is hearsay.

Note. SWARM is the only axis where the live board reports a SEPARATED leader (n=37, acc=0.384). The other 13 model-comparison axes are TIE or UNTRIED.

What's NOT here

This is a measurement axis, not a certification axis. The board never claims this model is "approved" or "compliant". A passing score is a public, signed record of what was measured, on which instrument, with which confidence interval — verifiable in your browser at /gspc-verify.

The numbers

n = the number of frozen item bank cells the model was scored against
accuracy = the deterministic grade (only on model-comparison axes)
separation = SEPARATED (one model is statistically the leader), TIE (multiple models tied), or UNTRIED (n too small)

Verify this axis

The live data is on /api/gspc. The canonical form is at /root.json. The witness receipts are at /signed/. The corrections ledger is at /api/corrections. The OTS anchor is at /api/state.