A society can be unanimous and unstable at the same time
When enough people publicly state preferences they do not privately hold, public opinion stops carrying information about private opinion. Nobody can tell how many others are pretending — including the people who are pretending.
Timur Kuran called this preference falsification. Its consequence is that an opinion climate can look immovable while being held up by nothing but each person's estimate of everyone else. A small change in perceived safety can then flip the whole distribution in weeks. Kuran's case study is 1989 in Eastern Europe, which almost no specialist predicted and almost everyone found obvious afterwards.
It means the apparent solidity of a consensus tells you nothing about its depth — and that “everyone seems to agree” is not evidence you are allowed to use. You are observing a display, and the display is the thing under dispute.
Granovetter's threshold model gets you most of the way there. Each person has a private number: how many others must move before they will. A population's distribution of thresholds, not its distribution of opinions, determines whether anything happens. Two populations with identical private beliefs can be perfectly stable and perfectly explosive, and no survey of beliefs will distinguish them.
The failure to foresee 1989 was not a failure of information gathering. In a threshold cascade, the outcome depends on the tail of a distribution nobody can observe and on the order in which people learn about each other. This is a different problem from an ordinary hard forecast — more data of the usual kind does not help.
It has the shape of an unfalsifiable theory.
Any surprise can be retrofitted as hidden preference revealed, and any non-surprise as preferences that were genuine all along. The theory earns its keep only where the private distribution is measured independently — list experiments and randomised-response surveys do this, and they typically find real gaps, but smaller and duller than the dramatic cases suggest.
Pre-registered measurements of private opinion that fail to predict subsequent public shifts better than public opinion does.