An Open-World Predictive Architecture Beside a Vision-Language-Action Model: Declared Rules, Prediction and Repair for Robot Manipulation

Vision-language-action models (VLAs) turn camera images and an instruction into robot motions. They keep no explicit model of the world: they cannot say what a step should achieve, notice that it did not, or be held to a rule. We place an open-world predictive architecture (OWPA) beside an unmodified VLA. A language model writes a declared world model of the task as readable rules; a deterministic engine runs it to plan and to predict the facts each step should produce; a monitor compares prediction and scene after every action chunk, re-instructs a stalled policy and repairs rules that prove wrong. We know of no earlier system that pairs a VLA with a world model written by a language model and repaired at run time. Tested with two open VLAs on LIBERO, in nominal scenes and in scenes changed during the episode, it gives three findings. The declared model is a reliable observer: it notices most failures early, names the fact that did not change, and costs nothing when the scene is as trained. Correction through language does not follow: better instructions, a repaired or perfect model, or a correct plan leave success unchanged, because the instruction selects among learned behaviours and does not compose new ones. A small fine-tune on demonstrations of the changed scenes is the only intervention that moved success. Detection and correction are separate problems; a declared world model solves the first and finds the data for the second.