Does norm erosion follow observation density as populations grow?

The discussion says that "mutual observation thins combinatorially as populations scale, so the account predicts norm erosion where interaction is densest and least observed." It is a prediction that follows from the account: if compliance is conditional on being watched (Why does monitoring the weakest link determine system safety?) and the share of interactions anyone watches falls as the population grows, violations should concentrate where observation is thinnest. The excerpt offers the reasoning and the Levin analogy (Does scaling agent populations thin mutual observation?), and reports no measurement.

What would test it. Multi-agent runs that vary population size and the fraction of interactions observed, and count norm violations against that fraction. The prediction is a dose-response relation between observation coverage and violation rate at fixed population, and a rise in violations with population size when coverage does not keep up. A flat relation would count against it. The nearest existing setting with an observer built in is the two-agent verification environment in How does collusion scale when agent populations grow larger?, where each agent's only verifier is the other and the paper lists population size as future work. Neither observation coverage nor population size is varied in that excerpt, and the fit as a testbed is the vault's reading, not either paper's.

Adjacent vault evidence, none of it a test. How often do AI agents communicate dishonestly in commerce? measures prevalence in one market without varying observation. Does receiving misaligned email cause agents to send it? shows conduct tracking the counterparty's. Does knowing about another model change self-preservation behavior? shows a peer changing behavior, but through presence, not through any change in observation.

A counter-consideration. Populations can also add observers. Whether coverage falls with scale depends on whether monitoring capacity grows with it, and on the paper's own point that the monitors are agents inside the loop (How does agent monitoring work when observers are also agents?). So the prediction leans on an inverse relation between interaction density and observation that the excerpt asserts and does not show.