Has anyone measured specification ambiguity as a predictor of correlated failure across model families? [D]

When models from different families are given the same underspecified task, they often fail in the same way rather than in independent ways. My question is about measurement, not explanation.

Has anyone put a number on the ambiguity of a task specification and then tested it as a predictor of how often independent solvers fail identically?

Specifically: does the relationship look like a smooth monotone increase, or is there a threshold — some level of ambiguity past which coincidence rate jumps sharply?

Looking for papers, metrics, or benchmarks where this was measured directly. Adjacent work is fine if you think it's close.