We describe an analysis of 3-GHz confusion-limited data from the Karl J. Jansky Very Large Array (VLA). We show that with minimal model assumptions, P(D), Bayesian and Markov-Chain Mone-Carlo (MCMC) methods can define the source count to levels some 10 times fainter than the conventional confusion limit. Our verification process includes a full realistic simulation that considers known information on source angular extent and clustering. It appears that careful analysis of the statistical properties of an image is more effective than counting individual objects.
