The authors investigate when a language model actually acquires task‑relevant information versus when it can express that information through its output layer. They argue that tracking correct token production conflates what the model’s hidden states encode with what its unembedding (readout) matrix can decode, because the readout continues to learn during pretraining. To isolate readout dynamics, they introduce parameter‑trajectory crosscoding: a single sparse dictionary is fitted to every token’s readout row at each training checkpoint, requiring only the checkpoint weights and no activation data. Applying this to Pythia’s pretraining reveals a narrow early window spanning 0.4 % to 1.1 % of total training steps in which the readout reorganizes and WordNet word categories become linearly decodable from its rows. Crucially, linear probes applied to the hidden states already decode task information at this stage, while the model’s own outputs still fail to do so. To test whether the timing gap depends on readout learning speed, they pretrain small models from scratch that differ only in the readout’s learning rate; increasing the rate advances readout development, narrows the gap, and leaves final loss largely unchanged. These findings suggest that early behavioral milestones in language model development may reflect the readout’s ability to express learned knowledge rather than the moment the knowledge is first encoded in the hidden representations.