We generate new score sets for SpamAssassin as follows:
- we freeze the codebase
- a team of contributors all download that version, and each runs the "mass-check" tool across their personal HandClassifiedCorpora
- they then upload the logs to a central server
- we run the GeneticAlgorithm over those logs
- it spits out an optimum score set
- we repeat step 2-5 once or twice more to handle the effects of the new scores on Bayes autolearning
Finally, we have a finished score set and we can release that.