A scaling law takes compute C and chooses the parameter count N and token count D. Each law assumes that N stays constant for the whole run without arguing for it.
GrowLab asks a different question: given compute C, choose the trajectory N(t). The answer is a whole curve rather than a single number.
The flat and grown arms share a FLOP budget and schedule shape, each at its own best learning rate, across multiple seeds. The first version of this result changed four things at once and ran one seed.
Both arms stop at the same FLOP budget. Lower is better.
This is the shape under test. Both arms must end at the same architecture.
Six hand-run points are too few for a search, so an autonomous agent proposes, runs and merges growth schedules against the objective.