Each collaborator is a point p in a cube of half-width 2.4 u. The hidden landscape is a sum of Gaussian idea pockets:
q(p) = Σ_k A_k · exp(−|p − c_k|² / 2σ²), σ = 0.85 u
Each fixed step (1/60 s) is a particle-swarm update with a local social term:
v ← w·v + κ·[c₁·r₁·(pbest_i − p) + γ·r₂·(pbest_h(i) − p)] + δ·u
p ← p + v·dt w = 0.72, c₁ = 0.9, κ = 0.25, |v| ≤ 3 u/s
h(i) is the colleague whose personal best member i can hear:
- Open plan — everyone hears everyone (one shared vision).
- Round table — only the k seats on either side (k = ring reach).
- Hub and spokes — a hub (member 0) hears everyone; the others hear only the hub.
- Hallway chats — whoever is within 1.1 u right now, so the network rewires itself as people move.
Sparse networks spread a good idea more slowly, which keeps several candidate ideas alive longer (more exploration, slower convergence). Dense networks converge fast but can lock onto a mediocre pocket (groupthink). The sweet spot depends on diversity δ and on how rugged the landscape is.
Checks. (1) With r₁ = r₂ = ½ a lone member obeys a linear recursion whose closed form (roots of λ² − (1 + w − dt·k)λ + w = 0) is compared with the live step function. (2) A round table of three with reach 1 is the same network as open plan, so both must give identical trajectories.
Honest simplifications. This is a stylised optimisation model, not a measurement of real teams. Parameters (w, c₁, κ, σ) are illustrative, ideas are one scalar quality, the landscape is static, and members never tire or argue. The showdown ranks networks only for this model and these settings. The 3D scene is a visual aid: the maths runs in the shared idea-space coordinates.