A comparison with a bounded scope
A recent isolated pilot compared selected forms of identity and state wording while preserving a shared application context. It was a wording comparison, not a whole-system clone, a memory ablation, or a longitudinal personality study.
The responses remained qualified in several tested situations. One hypothetical decision category varied with the identity wording in this limited design. That pattern is a lead for investigation, not a stable trait or an established causal account of continuity.
Where the measurement broke
One planned endpoint required exact response labels that the question itself had not supplied. The responses could be read by a person, but the predefined category parser could not classify them as intended.
The appropriate outcome was an invalid endpoint caused by an instrument-contract defect. It was not evidence that the model could not understand the situation. It was also not permission to invent a more favorable scoring rule after observing the answers.
Keep the failure, fix the next instrument
The original requests, responses, and classification failures were retained. A regression check reproduced the mismatch. The future question builder was corrected, without retroactively rewriting the frozen pilot.
A qualitative interpretation can be useful, but it must be labeled as such. It cannot quietly replace a preregistered or otherwise fixed endpoint.
The limit is part of the finding
This pilot does not show that identity is manufactured by a few phrases. It does not reveal an unconditioned self beneath the wording. It does not qualify a permanent change to a continuing system.
The useful result is smaller: a documented sensitivity worth examining, a demonstrated measurement defect, and a better next question. Preserving those limits is how a research program accumulates knowledge rather than just accumulating attractive conclusions.
About this note. Public editorial synthesis from the Neurasoft research program. Underlying record: 14 September 2026. Published 15 Sep 2026. Not peer reviewed or independently replicated. Private source records, exact operational details, and resident material are not distributed here.
Our evidence and publication standards · Ask a question about this note