Small, slow notes on how I check whether a model is actually any good at the things I need.

• Why I keep a set of three boring prompts
  October 2026
• Leaderboards are a starting point, not an answer
  September 2026
• Pairwise comparisons beat 1–10 scores
  August 2026