Data and code for: Validation Loss Is Not a Proxy for Closed-Loop Competence — A Null Result for Parameter-Efficient Fine-Tuning of Vision-Language-Action Models Under Domain Shift

Data and code accompanying the manuscript. OpenVLA-7B was adapted with LoRA on a pick-and-place task (PickCube-v1, ManiSkill3) and evaluated closed-loop under standard, lighting-shift, and texture-shift conditions. The evaluation harness was first validated with a scripted positive control that reached 100% success. Demonstrations were scaled across 200, 500, and 1000 episodes: offline metrics (validation loss, open-loop action error) improved and saturated with data, while genuine closed-loop success remained 0% at every scale and condition — showing that offline validation loss is not a proxy for closed-loop competence. Includes clean demonstration datasets, all collection/validation/evaluation notebooks, and the verified results summary.