In the first instance, you have a rating of 1992, with some built in "uncertainty" about how accurate it is. Your opponent has a rating of 1725 which may be a more accurate representation of his "true rating".

So, your 17 point penalty reflects two things:

1. You lost to someone "far" below you with a more certain estimate of their strength. This reinforces the notion that your "true" rating has not been properly tuned yet. So, there was a larger adjustment.

2. Their rating was a more accurate representation of their performance, so you "should have" been able to win if the 1992 rating was accurate, which the systems stipulates is not the case yet.

Your subsequent win of 19 points reflects the fact that the previous "rating correction" was too aggressive. Your 1975 rating was more accurate than your 1992 rating, based on the information known at the time of that previous event. But now your opponent was more accurately rated at 1760, and you won.

Your opponent was penalized 33 points because his gain from the last game was "too high". He lost to a more accurate rating from a lower rated opponent, so the penalty should not be as great.