Contribution
It generalised the 2007 result to a richer model and found a second route to the same benefit (stubborn priors). It also found the failure that comes from combining the two routes, and named the underlying good: transient diversity.
Key insight
Diversity of pursuit is valuable only on the way to agreement. Anything that keeps minority options alive also delays agreement. The right amount is enough to survive misleading early evidence and not so much that the evidence never wins.
Assumptions
- Two methods with objective success probabilities 0.5 and 0.499 (n. 12); agents hold beta priors with α, β drawn uniformly from an interval; myopic choice; Bayesian update on own and neighbours’ outcomes.
- Runs stop after 10,000 rounds; failing to agree by then counts as failure.
Key results
- Networks (Figures 3–4): cycle > wheel > complete; over all six-agent networks, success falls with density.
- Priors (Figure 6): as the maximum α, β rises towards 10,000 the order reverses; at extreme priors the complete network is best.
- Variance over time (Figure 7): diversity of actions persists much longer under extreme priors and in sparse networks.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Limiting information or making priors extreme each maintains diversity | strong | simulations, Figures 3–7 |
| C2 | Both together make diversity too stable and learning fails | strong | Figure 6 at high priors for cycle and wheel |
| C3 | Dense communication contributed to the delay in accepting the bacterial cause of ulcers | weak | historical narrative read through the model |
| C4 | Rationality in science is a property of groups more than individuals | weak | interpretive, after Hull |
Method
Agent-based simulation of a two-armed bandit with beta priors, on the cycle, wheel and complete graphs and all networks of up to six agents.
Concepts
- transient diversity — variety in what members pursue that lasts long enough to prevent premature abandonment, and then gives way to consensus.
Connections
Extends LIT-724. It sides with Kitcher and Strevens, against Kuhn, in showing diversity without diverse inductive standards, while adding a cost of diversity their models lack.
Bearing on the record
Source of THEORY-148. Rosenstock et al. (LIT-733) find this the robust result of the literature. For the owner’s essay it supplies a formal case with the shape of “protective lock-in” (§5.4), where a protective regime comes to sustain itself: the mechanisms that protect a minority commitment become, past a point, the mechanisms that prevent the group from ever settling.
Limitations
The same narrow model as 2007, at a success gap of 0.001. The ulcer history is an illustration, not a test.
Open questions
How to tell, in a real community, whether diversity is transient or entrenched before the outcome is known.