Why does an appended "abs"-block change earlier results?

I've got a Simulink model, simply a filtered step given as reference onto a control loop with PI control. There are Delays both in the plant as in the measurement feedback. As the simulation results oscillate heavily (wrong) I take ode15s as Solver. The result is good.

Now I add an "abs"-block to the system output, i.e. not inside the loop. The result BEFORE this block still looks good, but different from before. How can this happen, which one is trustworthy?

[Jochen Schuettler]

I use R2015a.

[Swarooph]

This seems like a solver related issue. Adding an Abs block maybe triggering extra zero crossing events which may affect how the model is solved as a whole compared to just that one block. Without looking at the actual model, this is just a guess of course. If you are new to the concept of zero crossing, take a look at this example and draw parallels with your application.