OFFICIAL MESSYTABLE PROTOCOL — FIXED100-SCENE SAMPLE 100 scenes from 100 distinct camera setups, 900 photos, 3,600 image pairs. 21,706 annotated appearances, 2,484,124 object-pair candidates, 78,335 true object pairs. Difficulty mix 30 easy / 50 medium / 20 hard. Selection fixed by SHA256 order before predictions. Metric values in percent Model | population | AP | FPR95 (lower is better) | IPAA100 asnet | published full test | 52.40 | 20.90 | 17.00 asnet | our 100-scene sample | 50.61 | 22.50 | 16.89 vit | published full test | 82.80 | 4.90 | 43.50 vit | our 100-scene sample | 81.29 | 5.46 | 40.31 AP measures how well true object pairs rank above false pairs across thresholds. FPR95 measures the fraction of negative object pairs accepted at 95% true-pair recall. IPAA100 measures the fraction of camera-view pairs for which every object association and absence decision is correct under the authors matching procedure. Original author metric functions and shipped Linux KMSolver ran in Python 3.7 Docker. Original authored crop/dataset and model definitions ran on modern Mac PyTorch/MPS. CPU and batch-size checks passed. Native released weights loaded strictly. No training. This sample supports the ranking and approximate scale of the papers results. Exact published full-test numbers remain unverified. The prior 100-pair fixed-budget recall experiment measures a different population and metric. Full official test includes 1,861 scenes and 66,996 image pairs. The full run was paused in favor of this requested sample. Sources https://arxiv.org/html/2007.14878 https://ijcai.org/proceedings/2023/149