Portability of a Disk-Anchored Agent Identity Across Foundation Models: A Blind, Internally Pre-Sealed A/B Comparison

Abstract

Large-language-model agents increasingly carry an identity — role, history, evolving state — anchored in external, persisted files rather than model weights. We examine whether such an identity can be operated with a non-native GPT-family engine using a four-experiment, blind A/B comparison. Honest scope. The arms differed not only in foundation model but also in reasoning effort and network-denial mechanism, so the study estimates a bundled arm contrast and does not isolate a causal foundation-model effect. On the deterministic axis, the intended historical anchor was redacted but intermediate/raw numeric residues remained; a residual anchor pointed the wrong way and was not followed, and an independent re-derivation reproduced the result without that anchor — evidence against that specific anchor mechanism, not proof the original run derived its result anchor-free. Judgment axis. The four-domain, eight-task decision rule was internally SHA-256-sealed before those runs and returned its label “Category A”: median difference +0.035, sign test 3–3 (two-sided binomial p=1.000), zero fabrication vetoes. We do not claim statistical equivalence: failure to reject a directional sign test is not an equivalence test, −0.650 and +1.594 lie outside the ±0.5 band, and no TOST or confidence-interval equivalence test was performed. The seal was internal, without external registration or trusted time-stamping. These are bounded descriptive observations from eight tasks in one ecosystem, under a confounded single Claude-family/GPT-family comparison. AI-transparency note. The experimental apparatus was operated substantially by AI agents under human direction (the human authors held go/no-go authority throughout); the draft was written by an AI writing agent and reviewed against the source ledgers by the human authors, and independently jury-reviewed cross-family (blind; the producing model families excluded from the jury). Foundation models are the study's subject and instrument, not tools of authorship.