Stefan contributed to this discussion with the presentation: 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝗶𝗻𝗴 𝗧𝗿𝗲𝗲 𝗜𝗻𝘀𝘁𝗮𝗻𝗰𝗲 𝗦𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗔𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀: 𝗔 𝗖𝗼𝗺𝗽𝗿𝗲𝗵𝗲𝗻𝘀𝗶𝘃𝗲 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗼𝗳 𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗔𝗰𝗿𝗼𝘀𝘀 𝗗𝗶𝘃𝗲𝗿𝘀𝗲 𝗟𝗶𝗗𝗔𝗥 𝗗𝗮𝘁𝗮𝘀𝗲𝘁𝘀 𝗮𝗻𝗱 𝗙𝗼𝗿𝗲𝘀𝘁 𝗖𝗼𝗻𝗱𝗶𝘁𝗶𝗼𝗻𝘀 Our main message was simple: as algorithms become more powerful, the benchmark itself becomes part of the method. We evaluated 12 open-source tree instance segmentation methods across 55 plots from 11 public datasets covering ULS, TLS and PLS data. We presented a benchmarking framework based on qualitative descriptors aligned with stand descriptions used in operational forestry. For tree-level point cloud analysis, this connects directly to the broader ForestSAT discussion: if close-range LiDAR for forest inventory, biomass assessment and change detection is to become operational in forestry, we need more than better algorithms. We need a community-based and structure-aware benchmark framework with practice-oriented descriptors, standardized pooling rules and clear support thresholds. The bottleneck is no longer only algorithms. It is also benchmark design.