In this work package, techniques rooted in computational information geometry will be brought to bear on the core statistical science problem of handling model uncertainty, while (Obj. 2) exploring their promise, and that of further challenging extensions of CIG, in graphical models and discrete data contexts. The operational tool to be developed is a new goodness-of-fit procedure for sparse multinomials. It will be used to determine which models – within CIG’s universal space – are data-supported. Deployed in concert with CIG’s ways to ‘think outside the model’ (reviewed above), this will allow the user to determine the effects on inference of moving between alternative plausible models.