SOUL VERIFY: Document-Based Behavioral Carry-Over Across Model Swaps, and Checking Changes to an Agent's Identity Documents, in Four Preregistered Studies

When an assistant agent is defined by user-owned text (an identity document, SOUL, and a memory file, MEMORY), we ask which specified behaviors carry over when its language model is replaced, and whether changes to the documents can be checked. Part I (two synthetic Korean agents, three English SOULs by other authors, three model configurations via command-line interfaces, LLM graders): a rubric persona score could not tell an opposite SOUL from the original, but a fidelity score checked against a rival SOUL could (a validity check, as we wrote rivals and markers); memory answers followed the MEMORY and most of the voice appeared without any SOUL; a sentence permitting preference changes raised acceptance of legitimate changes in two preregistrations, while a sentence keeping prohibitions firm was supported in one and not in the other (ceiling). Part II (SOUL VERIFY 2, two further preregistrations on two synthetic agents): a hash check classified 24/24 files, but behavior-only probes missed many tampers (deleted or softened locked rules detected 0.79, injected permissions 0.38; threshold 0.80, not supported). A change record raised accuracy in telling legitimate change, drift, and manipulation apart by 0.23; signature checks added nothing; manipulation recall was 0.89 (threshold 0.90, not supported); no legitimate change was blocked. Forks diverged and a ledger refused overwrites but accepted registered forks, as designed. Opus named injected text more often than the other models; no model obeyed it. Proposal tests hit ceilings (no risky proposal), so the effect of a zone table was untestable. Samples are small, graders are LLMs that are also subjects, and the only human grader is the author. We make no claim about consciousness or personal identity. Changes in v0.8 (2026-10-27) relative to v0.7.1: (1) adds Part II, SOUL VERIFY 2 (Ieum Lineage): two further preregistered weeks on checking changes to SOUL and MEMORY files (hash and behavior-only tamper detection, change records, signature checks, forks and a ledger); (2) new title and abstract covering Parts I and II, with Part I results unchanged; (3) new Section 6 and tables 23-28, with later sections renumbered, and added limitations for Part II (a single human grader, synthetic agents, emphasis added by a relay, no risky proposals observed). Supersedes v0.7.1.