Currently there's an open issue related to the probabilistic calculator using the sample based approach. Given the set of expected results and engine config, the related qa test fails in an non-deterministic way (for example in 10 runs it fails once).
Understanding the cause is not simple, the steps below are the ones I'm following to analyze the problem:
1) Produce end results using mean based approach with NUMBER_OF_SEISMICITY_HISTORIES = 100
2) Update qa test fixture, results can flux in a 5% range.
3) Run qa test with the sample based approach tracking down the following informations:
* Ground Motion Field Values
* Epsilon Values
* Loss ratios
We keep these intermediate results when the test fails.
4) We compare these intermediate results against a different
implementation (as provided by Vitor).