Some runs encounter a bug, where the computed expected value barely moves acording to new data, or oscillates between two (barely moving) values.

In order to fix that, detailed logs are needed, but currently Jenkins runs do not report enough details (or they are affected by rounding errors) so simulator does not reproduce the buggy behavior.

Make Jenkins logging precise and verbose enough, replicate the bug in simulator and fix it.
Probably rewrite the parts of code not readable enough for Vratko to spot the bug.