Study Tests Whether Task Rules Hold When the AI Model Is Swapped

A new paper in the Artificial Deterministic Intelligence series examines whether a bounded operational contract survives when the underlying inference model changes. Task state is kept outside the model weights, and the researchers do not assume smaller models match larger ones in capability. In synthetic first-party tests, a frozen governed package was moved between 8B and 35B models in both directions, with same-model restart controls, and all 120 confirmatory handoff trials kept the measured policy contract and task correctness. The study also compared governed scoped context, authorized metadata-filtered lexical retrieval, and an oracle ceiling, with every configuration completing 90 of 90 trials and conventional lexical retrieval showing lower retrieval-component time. The author states these results do not establish general model equivalence, production readiness, universal safety, or retrieval superiority, and the companion report includes non-acceptances, timeouts, and a non-confirming replication attempt.