Can you explain a bit more about this Jay? What are you hoping to see implemented in SmarterMail? The bayesian filtering method in place currently emulates a hidden markov model implementation, as only the end state is typically available to the product. Each message is scored against a series of tokens, and a probability of spam is then generated and applied to the message. This means that only the end result is visible (spam, or not spam) and none of the tokens are directly available to the system or end user. What information are you hoping to see leveraged in the antispam process?