Abstract
Conserving near-shore island endemics requires workflows that are robust to small, spatially clustered samples and that translate Species Distribution Model (SDM) into regulation-ready actions. We formalize a transferable SDM-to-action blueprint-(i) cluster-aware spatial holdout (leave-one-cluster-out, LOCO), (ii) conservative, high-specificity binarization paired with simple ecological filters, and (iii) explicit area-band uncertainty-and apply it to the Ulleungdo (Republic of Korea) endemic Scrophularia takesimensis. We combined 2008-2024 field records with a 5 m resolution MaxEnt model (linear-quadratic features; regularization RM = 1.40) using 28 unique presences versus 744 background points sampled within an accessible coastal belt (300 m from shore). Under LOCO, the model generalized well (AUC = 0.984 ± 0.014; partial AUC at specificity of at least 0.90 = 0.935; RelRMSE = 0.107) and mapped a narrow near-shore suitability belt with a continuous northern-northeastern core and fragmented southern-eastern satellites. To obtain a regulation-ready map, we converted continuous suitability to binary using a cutoff that achieved specificity of at least 0.98 under spatial holdout (threshold: 0.472; baseline: 300 m) and applied two ecological filters (retain areas within 90 m of shoreline; remove patches < 75 m(2)), yielding a CORE of 1.148 km(2) that captured 71.4% of recent records with zero leakage beyond the belt after post-processing. Accessible-mask sensitivity (masks of 300, 450, and 600 m) bounded the post-processed CORE to 0.930-1.593 km(2) (coverage: 0.607-0.789), which we carry forward as a planning area band. We translate these results into a tiered plan: protect the near-shore core, reconnect the fragmented southern and eastern stretches, and survey the highest-ranked coastal segments. Beyond this case, the blueprint generalizes to other small-n near-shore endemics, offering a transparent path from the SDM to policy while clarifying that, given static predictors, inferences concern present-day suitability rather than climate change forecasting.