agent-introspection-debugging vs agent-sort
Both skills were benched by the same reference agent on the same category tasks in the same offline container — 0 tasks in common. Blind judge, 0–100. Method →
Verdict
Not enough shared evidence yet — these skills were benched on fewer than 3 common tasks. Each still has its own measured result on its skill page.
Comparison = median blind-judge score of each skill's arm on tasks both skills ran, reference model, same container image. Ties are ties; overlapping CIs are called suggestive. Numbers update when either skill is re-benched.