Image recognition with indoor cameras is often hindered by insufficient illumination. This paper proposes a zero-shot low-light image enhancement method that requires no paired annotations or model fine-tuning. We treat a pretrained large-scale image editing model as a strong prior and formulate the low-light enhancement task as a composition of editable operators driven by task-specific prompts. To prevent the editing model from generating pseudo structures or over-imagined content, we design a physics-consistency scoring function based solely on the original and candidate-enhanced images, enabling unsupervised evaluation without relying on any additional perceptual networks or external data. The final enhanced image is obtained either by candidate selection or by weighted fusion according to the computed scores. We evaluated the proposed approach on public real-world low-light datasets and conducted comprehensive ablation studies. Experimental results demonstrate that, without any fine-tuning or supervision, our method effectively enhances low-light images and achieves comparable quality with fully-trained models.