The architecture scales with hardware, by commitment

The mission is efficiency against big-budget labs, so the architecture must have no scale dead-ends. Presets nano through one_b share one skeleton. RAM growth buys Engram rows (log-linear loss payoff, input side only). VRAM growth buys a bigger preset and batch. Machines buy data parallelism. Distillation enters when a teacher budget exists, teacher near 2.5x student size (arXiv:2311.07052).

Consequence: mechanisms whose benefit shrinks with scale are rejected even when they help today, and every small-preset A/B is read as a vote on the flagship.