Who Guards the Guardians? The Challenges of Evaluating Identifiability of Learned Representations — ThinkLLM