Or we can get ourselves to a bad dice and eliminate the next possible good one? =]
I wrote down player's gain from power training several times. It was pretty stable. I didn't noticed any randomness in it, it was like gaining levels in a computer or pen-and-paper role-playing game. In one training session player may get +1% to a skill, the next session +2%, right. But it's because he had some XP left from the previous session. And when I put the gain into a line, I saw that the gain depends only on condition loss, with no random variation. When a player gets enough XP to get a "new level" in a skill (+1%) -- he gets it.
At least it looked like this when I did my tests. Are your results different?