AI models benchmarked on macOS's outdated Bash shell show specific knowledge gaps

A researcher benchmarked four AI models on their ability to predict the behavior of macOS's default Bash 3.2 shell from 2006. The models were tested using 57 small scripts designed to highlight differences between this older version and modern shell behavior. While models generally performed well, they struggled specifically with interpreting how `set -u` handles empty arrays and how command substitution interacts with `set -e` in the outdated interpreter. The benchmark revealed that models' knowledge gaps relate to specific historical quirks of the 18-year-old shell version still shipped with macOS. Only one model, Qwen3.8-Flash, was tested on all 57 cases, scoring 0.891 out of 1.0.