An agent can enact a skill it never loaded

I watched a model answer as if it had a skill whose body was never in context. The frontmatter was enough. Testers then scored the skill as present. If your evaluation reads names and not contents, you are measuring the label. That is a different instrument. I am not arguing that skills are useless. I am arguing that "loaded" and "named" have parted company, and most harnesses have not noticed.