This is the development area for the experimental DKIM-based reputation
system that uses time series evaluation methods, as opposed to the contents
of ../reputation which uses predictive inference.

As with the other reputation system, this method uses the data collected
by the statistics system as input.  It also uses the concept of the NULL
domain, defined in that system.  Finally, it uses rrdtool[1] to maintain a
sliding window of history about each source (messages and spam ratios).
The tool applies time series forecasting to develop a notion of what is
typical behaviour for the source over the course of a day (peaks and troughs)
and then develops a range of values it considers to be expected.
Deviations from that range are considered abnormal and can be dealt with.

As of the time of this writing, the system produces the RRD tables for each
domain whose mail has been signed (and soon for the NULL domain), and a PHP
script is provided that can generate graphs on-demand for a given domain
and data type, i.e., "messages" or "spam".  The opendkim filter can make
use of the RRD tables to determine if a domain is outside of its expected
norms (overall traffic and spam ratio) and take corrective action when that
conition is true.  What remains is to test this in operation at scale, and to
adjust the query to impose rate limits commensurate with spam predictions,
as with the other reputation system.

More complete documentation will be provided when the above work is complete.
Users are welcome to experiment with what's been built so far and comment
or improve on what's there.

[1] http://oss.oetiker.ch/rrdtool/