Journal
· Niklas Nilsson
What AI-written code breaks, and why the metrics will not tell you
We measured four of our own codebases, two of which crossed from human to AI mid-life. Every metric we could extract described tempo. None described quality. What actually broke had to come from notes, and it had one thing in common: all of it compiled.
· Niklas Nilsson
Is this codebase worth saving? We ran the assessment two years late.
Nobody assessed the system before we restarted it. The client wanted to sell again and a rebuild was never on the table. So here are the four measurements, run in hindsight on 935 files and 64,000 lines — including the one where our own worry turned out to be wrong.
· Niklas Nilsson
The code was fine. The ground it stood on had rotted.
Three repositories behind one product went dark for up to 928 days. Waking the web admin took 24 lines of source. Waking the mobile app took 346 files. The difference is the whole lesson.