What these results do not say
They do not say what share of real AI slop AIV would catch. Disclosure is not a defect. A long function is not unused code. Closed and not merged is not slop. The 15,575 index was not an AIV run.
Measured limits
- On the added lines, a new function of 80 lines or more appears in about 9 percent of AI-disclosed patches and about 11 percent of merged controls. A length block would fail both.
- A new name that appears once in the diff appears in about 51 percent of disclosed patches and about 22 percent of merged controls. That is not unused code, because the caller can live in a file the pull request did not change.
- Four Airflow scheduler checks already exist and stay at warn: start_date=datetime.now(), missing catchup=, a sensor with no timeout=, and an empty execute(). This corpus is mostly core and provider diffs, not DAG files, so those rules barely appear here.
The small fixture suite remains 35 expected failures blocked and 15 expected passes clear. That is agreement with designed fixtures. It is not this full run.