Dark patches are indeed the worst case. The problem is that the i1iSis uses an i1 for the measurements. On a ~1500 patch target we typically see the worst 10% of patches having deviation of ~0.25 dE-2000, with the worst case being 0.5 - 0.6 dE-2000 if the measurements are driven by Argyll or MeasureTool and worst 10% of 0.3/worst case 0.6 - 0.8dE-2000 if i1Profiler is used to control the i1iSis.
Compared to X-Rite's older instruments, this isn't as good (Spectroscan: worst 10% <0.2, worst case <0.4; iCColor: worst 10% <0.2, worst case < 0.45).
The real problem is the occasional flyer where the measurement appears not to have be made in exactly the correct location. The most recent i1Profiler kicks out measurements that do not match the expected values. The MeasurementTolerance field in XRi1G2WorkflowSettings.ini controls the sensitivity. Unfortunately, i1Profiler won't give the data from measurements that are rejected, and from the best we can tell, the pass/fail criteria are based on assumed measurement colors for a given input (RGB or CMYK) value.
Our charts include a number of repeated color patches. These are used to determine orientation or page placement sensitivity of the printer and, relevant to the iSis, whether measurement errors occur. We process the files through our own profiling code which checks these patches. We have seeninstances where the iSis gives deviations of over 2 dE-2000 between two patches that should have the same color. Remeasuring showed they were identical, as did spot checks with a Spectroscan.
The problem appears to be that the measurement placement is slightly off, so the data are corrupted by a neighboring patch. Supporting this theory, the worst case tends to be between patches measured at the top corner of the chart compared to those at the opposite corner at the bottom. The increased measurement speed of i1Profiler over MeasureTool or Argyll makes matters worse. Without the ability to verify measurements independently of i1Profiler all you get is inaccurate profiles.
The measurement problems can be reduced by either using a slower measurement tool (e.g. Argyll or MeasureTool) or by modifying the target. Bumping the patch size up from the minimum of 6mm helps. Having a final, printed dimension of at least 6.3mm reduces measurement outliers to a rare occurrence (Target image size depends on how you print the charts. If you use ACPU, with its resizing, you need to make the charts accordingly larger). How you feed the charts into the iSis also matters. Having the chart centered in the measurement feed (and being less than maximum width) prevents the paper from bumping against the side of the slot.