Replication package: diagnostic evaluation of multimodal large language models on multi-image spatial visual question answering

Anonymous replication package for a manuscript under double-anonymous review. Contains experiment code (three-phase NIMES protocol, ablation studies, option-order ensemble mitigation, cross-benchmark validation on MuirBench, automated error taxonomy) and raw per-sample model outputs (JSON). The supplementary diagnostic model is anonymized as model-s; a domestic API endpoint is represented by the generic identifier doubao-pro-32k. Datasets: MMSI-Bench and MuirBench (public Hugging Face datasets; downloader script included). See README.md for reproduction steps. License CC BY 4.0.