Towards a grammar of Bayesian confirmation