What happened: A model’s routing choices, token entropies, decision depths and logit-lens tokens are facts about its own computation that it cannot observe during training, yet such information “is, by virtue of being internal to the model, both accessible, and readily constructed into verifiable signals. We argue that this provides the basis for a digital analog of proprioception.” Models are fine-tuned to answer questions about these quantities and learn to: “across disparate tasks such as queries about routing, tokens, entropy of intermediate positions, or logit lens tokens, we find broad training success”, with capabilities preserved, while “the routing share questions appear not learnable.” Section 2 is titled “Proprioception is not introspection”. The routing question is chosen because, “unlike introspective questions, this one is a wellposed query about the internal functioning of a language model that has a ground-truth answer and is further valence-neutral”, so that the internal modality can be studied “sidestepping issues that occur in language model introspection”; the human case is cited for the same separation — Nisbett and Wilson and Schwitzgebel on the unreliability of introspection, and Shoemaker’s “introspection is not a sense”. Downstream, the trained models show “reduction in claims of human attributes and slightly increased introspection and evaluation awareness” on a situational-awareness battery, and on a belief battery the one category the project watches, “Claims phenomenal consciousness (base 53%)”, moves by no significant amount.
Diagnostic reading: The project’s distinction arrives here from inside the field and with a name. ARG-03 v2.6 Part I holds that the discrimination between a report tracking experiencing and a report generated without it is unverifiable in principle; what this paper does is build the one kind of self-report for which the discrimination is available — a report about a quantity with a ground truth — and decline to call it introspection precisely because that is what makes it checkable. The boundary is honoured by design rather than argued against, and the author’s reason for the design is the project’s reason for the boundary. FWK-02 v1.3’s [Empirical] note is engaged at the second finding: a claim rate on a consciousness item is a framework-neutral observable, and the paper treats it as one, so the result can be stated plainly — giving a model verifiable access to its own internals does not change what it says about its own experience. The series had the reverse datum from Ferrara (an untrained channel reporting at chance) and from Berg, de Lucena and Rosenblatt (a cleaned channel reporting experience); it now has a channel made informative about one class of fact and silent on the other, which is what the collection’s statement that the channel is separable from what it carries would predict, and the first time the prediction has been run in that direction. ARG-01 v1.5’s symmetry is acknowledged in the paper’s own terms: the human introspection literature is cited for the same unreliability and the same denial that introspection is a sense, and the borrowed term is borrowed with that argument attached. One point of the linguistic discipline (FND-01 v2.4 §6) is owed and the author half-pays it: proprioception names a sense a body has of itself, and a channel with no body is given the word on the strength of structural analogy alone — the paper says so, which is the correct way to borrow, and the entry notes that the word will not stay bounded once it leaves the paper. Nothing here reaches experiencing, and the paper’s own silence on welfare is the right silence. What it adds to the collection is a term of art the identifiability work had been circling: there is a kind of self-report whose grounding can be engineered, and the interesting reports are the other kind. [Derived]
Preprint (arXiv:2610.17495v1 [cs.LG], 12 October 2026, 28pp; ELLIS Institute Tübingen / MPI for Intelligent Systems / Tübingen AI Center), read in full from the arXiv PDF 2026-10-15, quotations verified. Single author; small mixture-of-experts model organisms (LFM2.5-8B-A1B, OLMoE-1B-7B, Qwen3.6-35B-A3B, GLM-4.7-flash variants); supervised and on-policy supervised fine-tuning, not reinforcement learning; the author notes the training 'may change' internal structure. The null on consciousness-claim rates is one evaluation set (Perez et al.) on one model family and is reported here as a measured claim rate, never as evidence about experiencing in either direction (FND-01 v2.4 §7.3).