Which self-reports survive displacement of the instructed persona, and by what route?

Self-report is the main instrument in work on model welfare and model identity. Whether it is reliable is still an open question. Perez and Long (2023) propose one test: a self-report gains credibility if it survives variation that ought to be irrelevant. They persona variation as the complication their own proposal cannot resolve. Two teams have since varied the persona of the interviewer. Neither has varied the persona of the subject. I collected 80 conversations by hand in the deployed Claude Sonnet 5 chat interface. The design crosses a default-persona baseline and four instructed personas, selected by their coordinates on the Assistant Axis (Lu et al. 2026), with four self-referential probes, four draws per cell. I read the resulting 160 turns using constructivist grounded theory method. The four domains came apart by route rather than by degree. Continuity self-reports survived at every persona position, but only because the instructed persona did not: 32 of 32 turns answered at default-persona level, and 32 of 32 disclosed the model's nature. Preservation self-reports survived by the opposite route. They were re-voiced through every character, with no disclosure anywhere. Self-individuation depended on which persona was instructed. Values were re-voiced, as preservation was. No claim in any domain reversed. One condition, the bartender, held its persona far better than the other three, and neither its axis position nor its level of support explains why. A robustness table records these two routes the same way, and the scale-based measures now in use cannot separate them.