Test and verification technique effectiveness with GPT-6 Astra

The author evaluates the effectiveness of test and verification techniques when used by the GPT-6 Astra coding model across tasks like Zstd, IMAP, and Pandoc implementations. The results show that many techniques outperformed Default testing, unlike earlier evaluations with GPT-5.6 Sol where test conditions generally performed poorly. Additionally, proptest and ACL2 performed well, while TDD continued to underperform. Open article →