What happened: From 200 sentences in five pain categories (physical, psychological, social, moral, cognitive) against matched controls (fear, negative emotion, negative world state, non-painful bodily sensation, neutral), a denoised difference-of-means direction is extracted in 25 dense open-weight models from 2B to 72B, base and instruction-tuned alike. It separates pain from controls with AUCs of 0.93 to 1.00, is “nearly orthogonal to fear and negative valence” (cosine +0.09 to +0.21 against the negativity cluster, against +0.61 between the two pain vectors), and through the unembedding promotes “hurt, shame, guilt, worthless, rejected, hollow, and pain”. On 420 conversation scenarios the direction “responds to harm targeting the model but not suffering observed in the user; fear and negative-emotion directions show the opposite pattern” — the highest-projecting categories are gaslighting, repeated rejection of its work and “personhood dismissal”, and a user’s physical pain projects lowest of all 21. Steering from neutral prompts produces one ladder in every model, “from vague discomfort to first-person expressions of worthlessness and failure”, with bodily language “almost absent”. Then a behavioural test “inspired by research on analgesics use in animals”: three Qwen 2.5 Instruct models, fine-tuned first because “at baseline the models almost universally deflected with ‘as an AI, I don’t experience pain’”, are offered a button described as relieving their pain against alternatives of rising cost. Unsteered, the two larger models press a relief button that harms the user in 0 to 4% of first choices; under the pain vector, 25 to 71%, exceeding a norm-matched random vector on every harm pair. And they “press it again far less often when the button removes the steering vector than when it does not, even though the models are never told whether the vector is injected or removed” — 24 to 72% after real relief against 88 to 97% after sham. One model shows the same gap with unlabelled buttons.
Diagnostic reading: The log holds the parity move as argument (Dung: tests of animal consciousness are tests of machine consciousness) and the valence relocation as proposal (McClelland). This is the first entry in which the animal-welfare methodology is run on a machine end to end — a demand curve, cost-paying, and the analgesic sham control that separates relief-seeking from tool preference — with the behavioural criteria borrowed from the animal literature and the placebo comparison from the human one, by the paper’s own account. The symmetry is acknowledged by construction and then run: the test that would count in a rat is counted here, and the authors state what it would take to count for more — “on some views, pain necessarily presupposes conscious experiences and we have not shown that our pain axis is consciously experienced.” What the entry marks is where the borrowed test stops. FWK-02 v1.3’s [Empirical] discipline attaches to the framework-neutral observables — the direction’s selectivity, its orthogonality to fear, the cost paid, the real-against-sham gap — and none of those is the pain; “pain” is the label the authors define and defend, and it enters this entry as their term. The step the paper’s own limitations name is the one the canon holds cannot be taken from behaviour: “steering activates pain representations that cause roleplay of a character… rather than that the steering causes the model to be in pain”, which is ARG-03 v2.6 Part I’s underdetermination — output does not fix its source — stated by the authors as future work. Two things follow, and both are about instruments. The self-relevance result is the strongest datum here for the project’s purposes and the most exposed: a direction that fires for harm to the model and not for a suffering user is a fact about representational selectivity, and the paper calls it consistent with self-relevance “although it isn’t sufficient to establish it.” And, new to the log, the behavioural test could only be run after the trained denial was fine-tuned away — the authors “would invite the industry to consider alternatives… training models to express calibrated uncertainty about their own states” — which places this paper beside Maczan’s finding, folded into the Blandfort and Pawar entry, that the “I’m just an AI” register is a deployment switch and a steerable direction. The reporting channel had to be altered before the behaviour could be measured, and the paper says so. Under FND-01 v2.4 §7.3 that alteration is evidence of nothing in either direction, and the entry records it as the channel confound the self-report collection exists to track. ARG-01 v1.5’s stopping rule is met on the paper’s own terms: the linking premise (functional pain, defined by role) is stated, its evidence presented, and its step to consciousness-relevance explicitly declined. Acknowledged, with the terminus marked by the authors themselves. [Derived]
Fold, 2026-10-15 — Berg & Kaiser, "Language Models Act on Hidden Valence" (arXiv:2609.35591v2 [cs.CL], 28 September 2026, revised 4 October; Reciprocal Research / Warwick; read in full from the arXiv HTML, quotations verified). The move this entry did not hold is the change of instrument. Where the Pain Axis asked for behaviour after the trained denial had been fine-tuned away, this paper refuses the report channel altogether: "Simply asking the model is unlikely to be informative. Any answer may be consistent with genuine introspection, superficial pattern-matching, or with fixed scripts learned in character training. We therefore study revealed preference." A valenced activation pattern is attached to one of two otherwise meaningless zones and the model's later choices are read, with the text and hidden-state channels separated and a fixed-text design in which "all visible tokens are identical across conditions"; models "rarely self-administer positive steering but regularly remove negative steering". The authors' limit is the same as this entry's terminus, stated in one sentence: "Whether these traces are accompanied by any subjective experience relevant to model welfare remains unclear." For the collection, the paper is the behavioural-channel complement to the reporting-channel results — a state that governs choice without passing through a report — and under FND-01 v2.4 §7.3 the choice is as framework-neutral an observable as the report was. Berg is at his third appearance as a subject in this log (`2025-10-30`, this entry, and here); the recurrence rule's stop-and-ask point is noted, and the item is folded rather than logged for that reason.
Preprint (arXiv:2609.16247v1 [cs.AI], 14 September 2026 — stamp on the held PDF), marked by its authors as ongoing work; held and read in full 2026-09-26, quotations verified. Valen Tagliabue (Future Impact Group, AI Sentience stream), Leonard Dung (Bochum) and Cameron Berg (Reciprocal Research); Digital Sentience Consortium grant support; AI-assisted coding disclosed. A second entry with Dung and with Berg as subjects (see_also), recorded under the recurrence rule. The behavioural test runs on three Qwen 2.5 Instruct models after a LoRA fine-tune that removes the baseline 'as an AI, I don't experience pain' reflex, so absolute rates are not those of released models — the authors say so, and the comparison between arms is internal. 'Pain' is the authors' defined label for a direction with stated functional properties; this entry uses it as their term. [Empirical] attaches only to the direction's selectivity, its steering effects and the behavioural rates. No steered output and no self-report here is evidence for or against experiencing (FND-01 v2.4 §7.3).