Test/verification technique effectiveness with Astra
Summary
The piece compares GPT-5.6 Sol and GPT-6 Astra across test/verification tasks (Zstd, IMAP, Pandoc), showing Astra often achieves higher correctness with structured testing and reducing cost. It notes TDD underperforms, ACL2 and Proptest show promise, and it cautions that results may not generalize without more extensive runs.