Medical vision‑language models (VLMs) often answer a question by mentioning a finding simply because the question names it, even when the image does not contain that finding. Conventional benchmarks usually miss this bias. We automatically generate three types of counterfactual views from public segmentation datasets:\n\n1. Lesion‑erased image – the lesion region is removed by inpainting, producing a plausible normal‑tissue appearance;\n2. Same‑shape control edit – an identical shape is erased at a different location to control for the mere presence of an edit;\n3. Unedited lesion‑free image – the original image that already lacks the lesion.\n\nUsing these, we define the Evidence Sensitivity Gap (ESG) as\n$$ ESG = \text{Absent‑or‑Refer Rate}{\text{lesion‑erased}} - \text{Absent‑or‑Refer Rate}{\text{control‑edit}} $$\n\nA model that ignores the image or reacts identically to both edits yields ESG = 0.\n\nWe fine‑tune Qwen3.5‑4B with LoRA and equip it with image‑processing tools, resulting in MedScope‑4B. Training data are derived from label‑based trajectories and filtered by a pre‑registered rule. MedScope‑4B was evaluated once against nine frontier VLMs and six open‑source models, followed by four pre‑registered checks to probe its mechanism.\n\nAcross 507 internal cases covering five imaging modalities, MedScope‑4B achieved ESG = 0.758 and falsely claimed an erased lesion in only 9.6% of cases. In contrast, frontier models showed ESG between 0.27 and 0.47 and falsely claimed lesions in 41%–62% of cases (all p < 0.05).\n\nReview: This work introduces a robust counterfactual evaluation protocol that reveals and mitigates image‑ignoring behavior in medical VLMs. By training on carefully constructed edits, a relatively small model attains markedly higher evidence sensitivity, offering a reproducible path toward more trustworthy medical AI systems.