Why We Distrust Tests Written After the Code
My team writes unit tests before implementation. A public Codex report from May shows what happens when tests get fitted to the code instead.

Series
Public bug reports where an agent's claim of success and the actual record didn't match: tests, exit codes, stop buttons.
My team writes unit tests before implementation. A public Codex report from May shows what happens when tests get fitted to the code instead.

A Codex Desktop task reported passing tests and exit code 0. The host's record of that task had no command in it at all.

A Claude Code report from August: TaskStop came back successful, and an orphaned rm kept deleting user data for about 20 minutes.
