Much of what SAMT 2006 is about comes down to a single awkward fact: the evidence is never clean. A classifier looking at a video frame, a recommender weighing a user's history, a system guessing whether two annotations mean the same thing - all of them work from signals that are partial, noisy and sometimes contradictory. The honest response is not to pretend otherwise, but to reason in probabilities.

That is why the European Commission session's call for probabilistic and temporal modelling and approximate reasoning reads as more than a buzz-phrase. A probability is just a disciplined way of saying how much a piece of evidence should move your belief. Combine several weak, independent signals and a confident estimate can emerge from inputs that are individually unreliable - the same logic whether you are fusing low-level media features or any other stream of observations.

It is a pattern you meet far outside the lab. Markets that publish live odds, for instance, are a public example of many noisy signals being folded continuously into a single number: platforms such as betpas update their figures as fresh information arrives, exactly the incremental belief-revision that Bayesian models formalise. You do not have to care about the subject matter to find the mechanism instructive - it is a working illustration of evidence being turned into a calibrated estimate in real time.

The lesson for semantic media is the one the keynotes keep circling: meaning extracted from raw media is rarely certain, and systems that acknowledge their own uncertainty tend to be the ones you can actually trust. Reading the odds, in that sense, is good engineering.