Just to add a little to Gang's response...
Yes, it does make sense for the correlation to be very
close to zero.
Your main correlations are with an interaction regressor,
one that is basically the same as the seed time series
during one condition, and is basically the _negative_ of
the seed during another condition (and zero, otherwise).
Half of the non-zero time the correlation is basically
1.0, and the other half it is basically -1.0.
When using a fast event design, it might not be as close
to exactly +/-1 during those periods, but one would still
expect to have more-or-less balanced positive and negative
correlations over those parts of the time series (assuming
the positive and negative periods are equally weighted in
terms of the number of TRs).
So one would expect the correlation between the seed and
the interaction term to be close to 0.
- rick