Very brief outlines: I've had more on my plate here this last week than 
I'd anticipated. These are more geared towards reputation than trust, 
per se. These are on my wish list for my ideal reputation framework.

1. Contextualised reputation.

Let's say there are contextualised assertions of trust or reputation. 
eBay avoids these by simply having a "trustworthiness" rating; 
wikipedia, on the other hand, desparately needs something like this. I 
might trust statements from A on subject S (with some trust level t); 
and trust statements from B on subject R (with trust level u). The 
questions arise when looking at chained or inferred trust levels (A and 
B both talking about the same thing, or describing/rating aspects of 
the 
same entity, etc). If S and R are distinct, the problem is pretty 
simple 
since you can ignore assertions that don't have an appropriate 
contextualisation.

However, S and R may be two different taxons (either from the same or - 
quite possible in a semweb context - from different taxonomies). The 
question is, can we merge reputation or trust ratings that are 
contextualised with related or overlapping subject areas? And how might 
one go about actually establishing what those levels of "semantic 
overlap" are? To answer the last question, we might look at appraoches 
that use evidence: eg, relative classification levels using a corpus 
that's been classified using both taxonomies. Or, that information 
might 
be extracted as annotations on a cross-schema mapping (I'm thinking of 
SKOS-style cross-thesaurus work, for instance).

I think this is an area most amenable to progress.


2. For reputations: differentiation between knowledge and opinion (or 
generic and specific statements).

Example: Hofstader and Penrose might both be considered knowlegable in 
the field of AI, in particular when talking about whether conciousness 
is algorithmically realisable. However, since their conclusions appear 
to be diametrically opposed, there's clearly more going on that a 
reputation system might wish to deal with. Formal criticism is where 
this really crops up: subject experts might rate each other's generic 
experience in a subject highly, but take issue with particular 
assertions, claims, published papers, etc.


3. For reputation systems: handling feedback loops. In real-world trust 
systems, small cliques of mutually-regarding entities abound (eg, PGP 
web-of-trust). For reputation systems with inferred reputations, such 
loops pose problems for evaluation within many formal frameworks. This 
is more the casewhere trust is measured on a sliding scale rather than 
"trusted/not-trusted" where one might look at simpler mechanisms, eg, 
"hop count" in a PGP scenario.


This is another area where there's plenty of scope for real progress.


4. Combining 2. and 3. above: an agent's opinion about 
its own self-reputation. Most formal reputation systems tend to start 
with an agent that totally self-trusts. But in the real world, we are 
capable of recognising that an expert in our field may be more 
knowlegeable than us. We might reexamine our own opinions in the light 
of an expert's assertions. It is not unheard of to hear utterances 
like, 
"the more I spoke to him, the more I realised that I knew less about 
[foo] than I thought I did". I don't think this is a hard-and-fast 
requirement, because it raises all sorts of issues (if you are not an 
expert in a subject S, how do you recognise those with superior 
expertise from the ratings and recommendations given by others? How do 
you select which of those others that you trust, given that your 
expertise is low? Are there valid external "oracles" of expertise - eg, 
citeseer?)