Transparency Helps Reveal When Language Models Learn Meaning