I'd have thought that if you kept the same validation you'd risk over fitting.

Clearly that does make it hard to measure. I'd think you'd want "equivalent" validation (like changing the SATs every year), though I imagine that's not really a meaningful concept.