Here is a list:
The papers with incorrect statements are: F. G. Cozman, Computing posterior upper expectations, International Journal of Approximate Reasoning, Vol. 24, pp. 191-205, 2000; F. G. Cozman, Calculation of Posterior Bounds Given Convex Sets of Prior Probability Measures and Likelihood Functions, Journal of Computational and Graphical Statistics, Vol. 8(4), pp. 824-838, 1999; F. G. Cozman, Computing Posterior Upper Expectations, First International Symposium on Imprecise Probabilities and Their Applications (ISIPTA), pp. 131-140, Ghent, Belgium, June/July, 1999.
F. G. Cozman, Computing lower expectations with Kuznetsov's independence condition, Third International Symposium on Imprecise Probabilities and Their Applications, pp. 177-187, Carleton Scientific, 2003.
gives a number of constraints that must be satisfied by a Kuznetsov extension; the theorem claims that the list is exhaustive, but the proof only contains necessary conditions, not sufficient ones. Indeed the Kuznetsov extension requires more constraints. Besides, the proof of Theorem 3 is used to prove weak union of Kuznetsov independence, which may lead to problems there too; it is an open question whether weak union holds or not.
J. C. F. da Rocha, F. G. Cozman, Inference with separately specified sets of probabilities in credal networks, XVIII Conference on Uncertainty in Artificial Intelligence, editors, pp. 430-437, Morgan Kaufmann, San Francisco, California, 2002.
has several problems. First, the result on complexity of inferences for credal networks is correct, but its proof is not (the result was latter proved correctly by Cassio Polpo de Campos and appeared in IJCAI 2005). Second, the inference algorithm is presented so tersely that several key steps are not given, and these steps are actually the ones that have the most complexity. Third, the theorem on why separately specified sets decompose the inference is incomplete; the decomposition happens as stated but not in all cases. And finally, the connection with the 2U algorithm is correct, but it is so complicated to write that you should use the 2U algorithm, not the variable elimination version sketched in this paper. Well, I hope I never write anything else with so many mistakes. In any case, everything in this paper has been superseded by papers at UAI2003, UAI2004 and IJCAI2005.
D. Kikuti, F. G. Cozman, C. P. de Campos. Partially ordered preferences in decision trees: computing strategies with imprecision in probabilities, IJCAI Workshop on Advances in Preference Handling, Edinburgh, United Kingdom, 2005.
contains a few problems with examples and calculations; an errata has been produced with corrected versions.
Fabio G. Cozman. Sets of probability distributions, independence, and convexity, Synthese, 186(2):577-600, 2012.
defines epistemic irrelevance as equality of conditional and unconditional lower expectations, where the conditional expectations are conditioned in every event B that contains values of Y. However, epistemic irrelevance only requires equality given events Y=y for every possible value of Y. Likewise, confirmational irrelevance must be defined only given events Y=y (in this case the examples showing failure of Decomposition and Weak Union hold; they are not valid for the definition that uses conditioning on every B containing values of Y). Additionally, the paper mentions that regular independence fails Decomposition and Weak Union, but the counterexamples it mentions are not appropriate for this.