Methods
The full methods, including the exact prompt, image preparation and statistical analysis, will be published here alongside the arXiv paper in preparation.
In brief
Each of 790 camera-trap events (79 bird and mammal species, ten events each, from five GBIF-mediated datasets in Belgium, Ecuador and Luxembourg) was one test case. Each system's answer was compared with an expert-checked species label.
Species accuracy counts exact species identifications across all 790 events. The benchmark score (Adjusted TIS) credits the highest correct taxonomic rank (species 5, genus 3, family 2, order 1, class 0.5) and subtracts 5 points when any stated rank contradicts the reference, so declining to answer scores better than a confident mistake.