Field report: 1000 real pull requests scanned, 0 hallucinated packages confirmed

We run a security scanner on public pull requests and publish what it finds — including what it gets wrong. This is the weekly number.

What the field looks like this week

Zero is the headline, and it is not good news for us.

A lot has been written about AI assistants inventing package names. On 1000 real pull requests we have confirmed none. The 10 candidates that came back 404 were reviewed by hand and every one was *our* false positive — sibling packages of a monorepo, and configuration keys read as package names.

We publish this because it partly undercuts one of our own arguments. The laboratory studies that report high hallucination rates measure what a model produces when asked to solve a task. That is a different population from what survives review and reaches a pull request, and the difference matters if you are deciding where to spend attention.

What we got wrong, and fixed

80 new correction(s) to the detector this week. Each one is pinned by a regression test built from the exact line we misread, in a real public repository, with the pull request cited.

That brings the public record to 80 corrections across 20 cited repositories. The number is counted from the test file, not written by hand — delete a test and it drops by itself.

Recent sources: a/b#1.

Check any of this yourself

• The raw record: https://api.magsolutionsai.com/quality
• The live field figures: https://api.magsolutionsai.com/measurement
• The corpus behind both, with every cited pull request: https://github.com/MagSolutionsAI/MagSolutionsAI.github.io/tree/main/evidence
• Written up in full: https://magsolutionsai.com/quality.html

Generated on 2026-09-18 from measurements taken on public repositories. Every figure above is resolved against the endpoints linked here before publishing, and re-checked weekly afterwards — if one stops being true, this article gets a dated correction at the top.

This report was generated from measurements, not written by hand: every figure above is resolved against our public endpoints before publishing and re-checked once a week afterwards. If one stops being true, a dated correction appears at the top of this page. How often we are wrong →