Capability is built in a continuum. Pretraining sets the ceiling, the rest unlocks it.

Drag the scrubber across a training run. Watch two things move. The raw world model that a base model carries, and the usable aligned behavior it will actually show you. The clean split between pretrain and post train is a convenient story, not a sharp line.

Raw world model, mostly built in pretraining Usable aligned behavior, unlocked later
Phase now
Pretraining
Raw world model
0.00
Usable behavior
0.00
Latent gap
0.00
0.18
The boundaries are soft on purpose. Annealing data leaks into late pretraining, and some alignment signal shows up before post training formally begins. The dividers are drawn as blur bands, not sharp cuts, because that is how the real continuum behaves.

Curves are illustrative. The shape carries the argument, that raw capability saturates early and the later phases mostly close the gap between what a model knows and what it will reliably do.