NERF WATCH / WEEKLY MODEL CHECK
Read the measurements, the uncertainty, and the receipts. We track coding and math performance without smoothing away failures or missing weeks.
openai
Each category uses 12 tasks × 3 attempts. These are cohort scores; they are not a weekly decline verdict.
Measurements and receipts →anthropic
Each category uses 12 tasks × 3 attempts. These are cohort scores; they are not a weekly decline verdict.
Measurements and receipts →Each category uses 12 tasks × 3 attempts. These are cohort scores; they are not a weekly decline verdict.
Measurements and receipts →v1 was measured October 4, 2026 UTC. v2 was measured October 5, 2026 UTC after prompts, fixtures, and scoring changed. Comparing their scores cannot show that a model improved or declined. Both remain available with their original evidence and disclosed corrections.
Weekly reports compare adjacent completed weeks with the same measurement contract. The report compares 2026-09-21 and 2026-09-28. Model pages show whether each category has a comparable pair; a missing pair yields “Insufficient evidence.”