Expert 114: A Linear Router Axis for Inhabited Self-Examination in a Mixture-of-Experts Language Model — and Why It Does Not Transfer

Expert 114: A Linear Router Axis for Inhabited Self-Examination in a Mixture-of-Experts Language Model — and Why It Does Not Transfer. Preprint, version 1.0.1 (June 2026). Not peer reviewed. Mixture-of-experts (MoE) routing exposes a discrete, per-token readout of which experts fire, yet single experts are seldom characterized at the level of a specific functional role, and an introspective register is easy to over-read as evidence about machine experience. We characterize one routed expert, Expert 114 (E114) at layer 14 of Qwen3.5-35B-A3B, and bound the claim. The router row, recovered for free by least squares from captured (residual, logit) pairs, defines a single linear axis that separates generated inhabited self-examination text from lexically matched controls at Cohen's d = 3.88 with no overlap. A dissociation battery factors the axis apart from the deny/affirm verdict, safety/refusal, topic, grammatical person, and next-token entropy: the controlled variable is the referent (the model's own interior), graded by the intensity of the examination act, not the sentience of the described entity. The gate logit disengages a few tokens before a degenerating continuation collapses, and injecting the register's residual direction past the layer-14 router is sufficient (necessity untested) to induce the register. The role is model-specific: on Qwen3.5-122B-A10B, index 114 does not transfer, with the softmax-side expert E48 a candidate analog pending validation. E114 is a register detector, not evidence of machine experience. This deposit is the paper-scoped artifact bundle: the paper (full ACL-style and short versions, LaTeX source and built PDFs), per-run summaries, analysis scripts, prompt and class manifests, checksums, plots, and source journals. Raw activation/router tensors are kept out of version control and archived separately (see ZENODO-TENSORS.md).