Trying to open the transformer inside out — studying transformers trained on Markov processes. Detection of non-linear truth representations in low-dimensional subspaces of the residual stream.