Continual Learning at Fixed Update Bandwidth: Budgeting Quantization-Cell Exits in Deployed Models

A deployed language model is a low-bit quantized integer model, replicated many times over, so every continual update must be serialized and shipped to every replica. This is a recurring cost that continual-learning research does not measure, and it scales with how often a model adapts rather than with how much it learns. We observe that under a \emph{frozen} deployment grid a group-wise quantizer partitions weight space into cells, and that moving a weight inside its cell leaves the shipped integer model bit-identical. The update that must actually be transmitted is therefore exactly the set of changed codes, which makes update bandwidth something a regularizer can act on directly. We propose Cell-TR, which budgets cell exits during continual fine-tuning and is evaluated by the size of the serialized code patch rather than by a proxy. Conventional regularizers do not reach this regime at any strength we tested: sweeping the EWC coefficient across three orders of magnitude saturates well short of it, and replay's payload is nearly invariant to its ratio. Cell-TR instead reaches 3.71~MB at accuracy .879, $14.0\times$ smaller than sequential fine-tuning with no detected loss of accuracy. Two ablations locate the effect in what the penalty is anchored to rather than in the shape of the barrier: widening the free region inside a cell only costs bytes, and against the current-bin-center rule of quantization-aware training, anchoring instead to the previously \emph{deployed} state is cheaper in 20 of 20 matched configurations. The advantage grows with update frequency, which is what makes it relevant to models expected to keep learning after deployment.