When I was working on the model of argumentation referred to above, Tony Hunter and Philippe Besnard started to look at paraconsistent logics. But these typically end up supporting conclusions that are somewhat counter intuitive. So they moved towards the preferred solution in the argumentation community of working with consistent subsets as the basis for an argument. In the case where we have on un-attacked argument for A and another against A then it is hard (not possible?) to find a rational way of preferring one or other outcome. Most models of argumentation allow a mechanism of undercutting, where a further argument can contradict a proposition in the support of an argument. That in turn can be attacked ... So without any notion of weighting of propositions, one is able to give a notion of preference of conclusions on the basis of preferring arguments where all their defeaters can themselves be attacked. In cases where ordinal or cardinal weights are allowed, then finer grained preferences can be supported. Going back to an earlier part of the discussion - it is possible to allow reinforcement between arguments if weights are supported. But you do need to account for any dependencies between the arguments (so there is no double counting). Our "probabilistic valuation" did just this (see section 5.3 of the paper Alexandros cited). In cases where you are unsure of the relationship between sources of evidence, the possibilistic approach of just weighting support for a proposition by its strongest argument (use of "max" for aggregating strengths of arguments) is appropriately cautious.
Background on Agorics:
The idea of software agents cooperating in an open market or "agora". Described by Mark Miller and Eric Drexler here: http://e-drexler.com/d/09/00/AgoricsPapers/agoricpapers.html Depicted by Greg Egan in his novel "Diaspora", exerpt here: http://gregegan.customer.netspace.net.au/DIASPORA/01/Orphanogenesis.html
Background on Argument: http://en.wikipedia.org/wiki/Argument
Let's start by supposing that an argument is a variety of persuasive message. If Bob trusts Alice though, Bob could be persuaded by simply recieving a claim from Alice. That is a kind of persuasive message, but it's not an argument. If Bob is insecure, then Bob's mind could be hacked and therefore changed. However, that's not an argument either. (The "Buffer Overflow Fallacy"?)
Possibly arguments are witnesses (or "certificates"), as used in computational complexity. Alice could spend exp-time to solve an instance of an NP-complete problem, then send a small witness to B, who can then spend poly-time to verify it. The witness would be an argument.
I'm not sure if that's a definition, but we have an overgeneral category (persuasive messages) that is, a superset of arguments, two subcategories of persuasive messages that are specifically excluded, and one subcategory that is specifically included, which seems like enough to go on with.
We know what witnesses to SAT problems look like - they look like satisfying assignments. That is, if Bob were considering a SAT problem, and Alice sent Bob a putative satisfying assignment, and Bob verified it, then Bob ought (rationally) to be convinced that the problem is satisfiable.
What do other kinds of witnesses look like? What about probabilistic computation? What if Alice and Bob may have different priors?