Report and Representation in a Second Model: A Registered Replication of Meaning Reorganization in Qwen3.8-27B

When a story’s ending shows the narrator’s belief was wrong, does the meaning of the belief’s own words move inside the model, and does its judgment go with that movement? This registered replication asks it of a differently built model, Qwen3.8-27B. Across the same ten stories, each ending came twice, once from outside the narrator and once from within; after each, the belief was judged true, false, or unknown, and the movement of its hidden-state representation was measured across 64 layers. The representation moved beyond wording after a revising ending, in every story and in both studies, in a single band of layers 18– 64 (p = 2/59,049 for endings from outside, p = 1/59,049 for endings from within), as in Gemma (Bruyn, 2026d). The one-word judgment followed the revising endings, FALSE in 20 of 20 versions in each study, but not the confirming ones, TRUE in only 4 of 20, leaning FALSE while uncertain. The registered rank correlation across stories was not significant in either study. H1 is split: supported for the revising endings, not for the judgment after the confirming ones. The design exists to find where the report and the representation part, and here that place is the confirmation, not the revision; an exploratory comparison found the movement larger after endings from within than from outside, layers 44–57, p = .033. Correction: in the reference list, the DOI of Bruyn (2026), “What next-token surprisal can and cannot see”, should read 10.5281/zenodo.23340596.