Meta-learning provides a robust framework to discern taxonomic carnivore agency from the analysis of tooth marks on bone: reassessing the role of felids as predators of Homo habilis.

Determining carnivore agency in taphonomic research is crucial for identifying site formation processes and carnivore-hominin interactions that influenced human evolution. Previous deep learning (DL) models classified the four principal carnivore agents affecting African hominins, but exhibited uneven performance due to unbalanced sample sizes. This study introduces a dual method based on few-shot supervised learning (FSSL) and model-agnostic meta-learning (MAML) as an alternative, achieving more consistent accuracy (FSSL: 81.54-83.56%; MAML: 82.56-85.13%), and significantly improving macro-average F1 scores. The best performing MAML model, Xception, reached 85.13% accuracy and an 84% F1 score, with taxon-specific F1 scores of 82% (crocodiles), 83% (hyenas), 88% (leopards) and 83% (lions), making the most precise classification of carnivore-made tooth marks to date. Applying FSSL-MAML ensemble models to Homo habilis specimens OH7 and OH65 from Olduvai Gorge confirms that leopards were preying on these hominins, as they had been earlier on australopithecines. Contrary to our expectations, these findings demonstrate that early Homo was still part of the prey spectrum, reinforcing the idea that the transition to dominant predator status occurred later in human evolution or penecontemporaneously to H. habilis through a different hominin taxon.