Zollman Lab — the original model

Bandit networks & transient diversity (Zollman 2010), his assumptions and his code. Two theories, and nobody knows how good either one is: the new one really pays off at 0.5, the old one at 0.5 − ε. Will the community find out?  ·  ← tutorial  ·  the LLM extension (earlier engine) →

Presets

Community

who sees whose results — Zollman's headline: sparser networks are slower but likelier to find the truth
community size — his Figure 3 runs 3 to 11, Figures 6–7 use 7

The problem

the new theory pays 0.5, the old one 0.5 − ε (his 0.5 vs 0.499) — small ε = a hard problem, easy to get wrong
how much data each scientist generates per round — the size of each "study" before everyone updates (his 1,000)

The scientists

Zollman's "extreme beliefs": every α and β is drawn from (0, 4 × strength] — his "maximum α/β" is 4 at strength 1, 10,000 at strength 2,500
every scientist holds a beta belief about EACH theory and works on the one that currently looks better. Untick to recover the earlier build's simplification: the old theory's rate is known exactly, only the new one is learned, and giving the new theory up is then permanent.

Simulation

how many independent communities to simulate — more runs, steadier percentages
the horizon at which the verdict is read: communities still divided count as "never settled". His paper says 10,000; his code defaults to 2,000; his Figure 6 fits 5,000
random seed — same seed + same settings reproduces the exact same result
Log sliders: ε = 10^(−3+x), trials = 10^x, prior strength = 10^x. Everything reruns client-side; nothing leaves this page.
Press Run.

The community

Outcomes at the horizon

Diversity trajectories (average share of scientists on the new theory, round by round, grouped by how the community ends up)

Diversity is a line sitting between 0 and 1: some scientists on each theory. It ends when a line reaches 1 (everyone on the new theory) or 0 (everyone on the old one). This is Zollman's Figure 7, drawn as a share rather than a variance.

When is the question settled? (share of communities that have reached their final verdict, by round)

Sweep: P(settle on the truth) by stubbornness × network (his Figure 6; uses current ε, trials, scientists, horizon, and the checkbox)

Zollman's Figure 6 (7 scientists, 5,000 rounds): the three networks coincide near max α/β 3,000, and past it the cycle falls away (0.77 at 10,000) while the complete graph holds (0.97). Dotted lines: share of communities still divided at the horizon.