Information theory explains the robustness of deep GP trees, with on average up to 83.3 percent of crossover run time disruptions failing to propagate to the root node, and so having no impact on fitness, leading to phenotypic convergence. Monte Carlo simulations of perturbations covering the whole tree demonstrate a model based on random synchronisation of the evaluation of the parent and child which cause parent and offspring evaluations to be identical. This predicts the effectiveness of fitness measurement grows slowly as O(log(n)) with number n of test cases. This geometric distribution model is tested on genetic programming symbolic regression.