In a small city there are say 100,000 people. If most of them carry weapons then the probability that a small number will be armed when having a bad day and losing their tempers with random strangers become real. In a sizeable population social dynamics are not what happens in a restaurant where each person can see the antagonist - the 50-person restaurant is not an adequate model of how people act in large scale. One of the factors affecting scale is that as a model is scaled up it has to model resource limits, that are typically not present in small scale situations. Large scale models also have to deal with closed-world assumptions that can often be left out in small models. People intuitively understand that these problems of scale are critical. The growth of events in social populations is not linear, it is typically exponential; including the comunication of fear and fear responses. As a simplistic counter-example to your recipe, following your restaurant shooting example, a surviving 1/3 of the diners being armed leave to find some people nearby to shoot. At a minimum the restaurant model needs to show how exponential growth of violence proceeds to be labelled real.
