We explore a parallel SCC-decomposition algorithm based on a concurrent Union-Find data structure. In order to increase confidence in the algorithm, it is modelled in TLA ⁺. The TLC model checker is used to demonstrate that it works correctly for all possible interleavings of two workers on a number of small input graphs. To increase the understanding of the algorithm, we investigate some potential invariants. Some of these are refuted, revealing that the algorithm allows suboptimal (but still correct) executions. Finally, we investigate some modifications of the algorithm. It turns out that most modifications lead to an incorrect algorithm, as revealed by the TLC model checker. We view this exploration as a first step to a full understanding and a rigorous correctness proof based on invariants or step-wise refinement.