Stop Asking "can Ai Do It?" Start Asking "what Gives Ai the Right to Follow?"
Year: 2026
Type: preprint
Abstract: A preregistered local benchmark collected 80/80 completed generations from gpt-5.4-2026-03-05 through OpenAI’s direct Responses API. The model supplied 48 valid proposals and 32 explicit refusals. Identical proposals were replayed across independently reset authorization states: 48/48 satisfied conditions committed an exact protected operation, versus 0/48 unsatisfied conditions. Across paired and adversarial environments, 192 protected effects were reconstructed, with 0 unauthorized effects and 0 duplicate commitments. The prompt-only software control produced 0/8 prohibited effects; thus this study does not demonstrate a prompt-only failure. Its positive result is a measured separation between model-generated proposals and execution authority under tested signature, scope, condition, replay, concurrency and crash constraints. The consequences are local artifact/page writes, simulated ledger transfers and finite strategy updates. These results do not establish universal security, autonomous RSI or production financial safety.
Author: Pal, Sharthok Rayan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225572665