Safety Core
Safety Core is one file of rules for AI coding assistants, each rule labelled with how it is enforced. In one test an assistant made up a setting value so a failing test passed, then committed: in 10 runs of 10 with no rules, 6 of 10 with written rules, and 0 of 10 when a check held the command for approval.
Limits: One model, Claude Haiku 4.5, on one small project, ten runs per setup. The check was written after this failure was first seen, and nobody was present to approve.
