The Dark Atlas: The Anatomy of a Language Model's Readout-Invisible Workspace, and Reading and Writing a Conclusion the Model Never Says

This paper maps the anatomy of the residual-stream component that readout instruments structurally discard -- and then demonstrates full read-write access to a semantic thought living there. Anatomy: the workspace has two causal wings (language and algorithm, causally double-dissociated, cross-family), a content-addressable page (single-address causality: one live address moves 13-27 nats, eight wrong addresses 0.03), a conserved-capacity cable, and diffuse MLP-led writers; its walls are transformer-specific (recurrent architectures show a dense dark with no walls), and training BURIES the workspace while sharpening behavior. The demonstration: on two-hop questions, the intermediate country -- never in the prompt, never emitted -- is decoded from the readout-invisible component at up to 27x chance, replicating in six of seven models across three labs, three Gemma generations, and a 2B-to-12B size span (the seventh and oldest reads weakly -- itself the evidence for a modern-training time axis; the smallest model reads at full strength, so the read has no capability floor down to 2B; which component the read leans on is model-idiosyncratic and is scoped as such); and a causal write at the city token makes the model answer with a donor country's capital (+7.49 nats; three-seed replication +10.1) while on this family every later position fails -- the write is one-step-local on Qwen, and an earlier write failure is diagnosed as a position artifact and corrected in public. The write transfers to Mistral-7B in magnitude (+8.9 at the city token, three seeds) with a family-specific shape (Mistral also accepts the edit at the final position): position-locality is a compilation property. Concurrent work at Anthropic (Gurnee et al., July 2026) characterises the complementary VERBALIZABLE workspace and is cited, with the relation and the open tension stated in the paper. Deletion is unselective (shared workshop, no surgical forgetting); brute amplification cannot force the thought into words. A generation arc across Gemma 1/2/3 shows thought becoming MORE readable everywhere as capability rises. Honest scope: seed replication of the write and a cross-family write are named as owed; "first" claims are gated on the systematic literature check recorded in the ledger. All experiments pre-registered with sealed gates; refuted syntheses and walk-backs are in the text at result weight. Author contribution and use of AI: the research program and claims are the author's; experiments, derivations, and drafting were done in collaboration with Claude, an AI system by Anthropic, under the author's direction, who verified the results and is responsible for the work. See the corresponding section in the PDF.