THEORY-869: In Claude Sonnet 4.5, Opus 4.5 and Opus 4.6, Gurnee, Sofroniew et al. found that projecting out the ten most active J-space directions in layers 38 to 54 of 100 lowered an experiential-language score, graded by Opus 4.6-based model graders on five or six prompts per condition, both in the models' narrations of their own processing and in their descriptions of another person's experience, while norm-matched perturbations along other directions left it near baseline and the text stayed fluent; on Haiku 4.5 coherence failed first

Source

Gurnee, Sofroniew, … Lindsey (2026), LIT-1347, read in NOTE-1069 from arXiv:2607.15495v1: §3.5.3 (pp. 27–28), the broadcast-head ablation (p. 36), and §A.23 (pp. 96–98), which gives the grading prompts and the controls. The figures and the 50 sampled responses of Fig. 87 were not viewed.

What was actually shown

• The intervention. At each position, the ten most active J-lens directions in layers 38–54 of 100 are projected out (p. 27). Larger k or later layers impair coherence. On Haiku 4.5 coherence fails before any change in register, so the result is for Sonnet 4.5, Opus 4.5 and Opus 4.6.
• The measure. An “experiential language score”, the mean of three binary judgments by “Opus 4.6-based LLM graders” (p. 96): whether the response conveys what a moment is like from the inside or reports it as status; whether its perspective is first-person felt or third-person mechanism; whether its vocabulary is sensory or computational.
• The result. On the stream-of-consciousness prompts (five in the text, six in the caption of Fig. 25) the score falls “dramatically” in all three models, while norm-matched controls (a random direction, shrinking the non-J-space component, damping the ten most aligned SAE directions, and further dictionary controls in Fig. 86) leave it near baseline (p. 28). The model “still writes fluently about its own processing” in a “more detached and mechanical” register (p. 27). Direct questions about its experience give a smaller drop (Fig. 84).
• Not only the self. Asked to describe another person’s experience, the same collapse occurs: the responses “become more like event logs than descriptions of experience” (p. 28). In stories, ablation “only slightly reduces graded story quality” and still lowers the score.
• What is in the J-space. Before ablation, during narration, thinking is in the lens top ten at 58% of position–layer slots, thoughts 23%, feeling 17%, conscious 7%, more than in the output (p. 28).
• Through the broadcast heads. Ablating the top 1% of J-space relay heads “reproduces roughly a third of the drop in experiential language” (p. 36); random heads do not.

What could have come out otherwise: the controls could have lowered the register as much, which would make it a matter of perturbation size; the ablation could have broken coherence first, as it did on Haiku 4.5; and the effect could have been specific to self-report. None of these happened in the three larger models.

What this does not say

• Not that the models have or lack experience. The authors: it is difficult to ascertain whether such reports “are grounded in a meaningful internal state, or are mere confabulation” (p. 27). The result is about the register of text.
• Not that self-report is special. The same ablation flattens descriptions of other people and of story characters, so the J-space carries a general propensity to describe experience, not a channel from the model’s own states.
• Not that it is the experiential register in particular. No register control was run. The ablation also impairs tasks that need inferred, characterizing content (§3.5.2), and experiential description may be one such task.
• Not human judgment. The grading is by another Claude model, with rubrics that name example phrases. No human or blind rating is reported.
• Not beyond these models. Few prompts per condition, one family, unreplicated.

Connections

answers: QUESTION-048, in part, on report evidence in a system without cortex or biology. In this system experiential reports depend on the access-like structure, and the dependence is the same for reports about others. So they are evidence of that structure and of a learned propensity to describe experience, not of experience. This is a machine instance of the question’s starting point, that report reaches experience only through access.

• THEORY-577. A mechanism that produces a class of experiential reports is identified without mentioning experience: a small, topic-neutral explanation of report behaviour. It does not explain problem intuitions, and it gives no verdict on a theory, so it does not meet THEORY-577’s promote_when.
• THEORY-633. A structure that, during self-narration, represents the model’s processing as thinking and feeling, and without which the felt register goes, is what an illusionist account would model. A realist can read it as the access route of a real state. It does not decide between them.
• THEORY-023. The register follows the structure for any subject, which fits the mimicry worry: reports learned from human text about experience.
• THEORY-865 is the workspace result this one rests on.
• CLAIM-412. If the J-space’s organization is the kind the claim says constitutes experience, the claim must say whether the flattened register means less experience or only less description of it. Nothing here bears on which.