I don't know how the Natural Breaks (Jenks) algorithm works in QGis graduated symbolism, and I've deliberately not researched it in order to ask this question, as my point is based on what seems to be intuitively wrong, rather than what may be technically correct (but not useful or expected).
So! If you have a range of values and constrain them to a few classes, Jenks does a nice job of splitting up the data:
But if you add further classes, Jenks just repeats the lowest value, often:
Since I don't know (yet - I will now research it) how Jenks works, it'd be remiss of me to say the algorithm is coded "wrongly", however the implementation is peculiar and unexpected. Possibly if it auto-constrained itself to the maximum number of independent classes? Or pooled identical-valued colour bands?