The data used in the demo is a simulated single-baseline observation
18 minutes long. There is one spectral window with 16 channels
covering 22.000-22.0075 GHz, with data for RR and LL (no cross-hands).
There are two fields observed alternately: a calibrator with flux
density 1 Jy (20s/scan) and a target with flux density 0.05 Jy
(40s/scan).  A single continuous time-dependent systematic phase delay
variation has been applied to both polarizations (emulating clock,
geometry, or atmospheric delay errors that are seen by both
polarizations). The magnitude of this phase delay variation was chosen
to introduce many cycles of time-dependent phase over the full
observation (thus fringe-fitting is needed).  Additionally, a
polarization-dependent (electronic) delay offset has also been applied
(43ns for RR, -86ns for LL).  These delays are small enough to
introduce less than 1 cycle of phase across the 8 MHz bandwidth (this
is all K can handle so far).

Three versions of this dataset exist:
  test0.ms   -  noise=0.001 Jy  (high SNR, to reveal systematic details)
  test1.ms   -  noise=0.1 Jy    (med SNR, more realistic)
  test2.ms   -  noise=0.5 Jy    (low SNR, to illustrate detection)