I'm surprised it continued to change after detail was lost. Seems like there would be some measurable property of a codec/decode process that could be quantified - the entropy? It would be a desirable property to be 'idem-potent', e.g. after one cycle the result settles.