Verbalizable Representations Emerge before Workspace Functions in Language Models

The J-Space provides a way to separate when internal representations become readable from when they begin to affect model behavior. We study its development across Pythia model scales and training checkpoints. Information from the current and earlier context becomes readable early, while latent intermediate representations emerge more gradually. J-Space ablations increasingly affect model performance later in training, indicating growing causal relevance. However, stronger functional properties remain weak. These results suggest that readability, causal relevance, and functional control mature along different timelines.