Language models often appear to become insensitive to how text is tokenized. Models trained with different tokenizers can reach similar language-modeling loss, and prior work shows that they recover word-level information from the raw subword segmentations (a phenomenon referred to as "detokenization"). This raises a natural question: once that information has been recovered, do the resulting hidden states become independent of the tokenizer that produced them? We study this using *tokenizer twins*, pairs of models matched in architecture, training data, and random seed but trained with different tokenizers. We learn small translators between their hidden states at positions where the two tokenizers agree, and then test the same mapping where their tokenization paths differ. Transfer works well at agreement positions, adding less than one nat of loss, but breaks sharply at disagreement positions, where the loss increases by more than seven nats. This gap appears from the first hidden layer and persists through the network. At the same time, word-identity probes can still recover much of the underlying content from these representations. In other words, language models can abstract beyond their individual subword tokens while still carrying a lasting imprint of how the text was tokenized.