A "for example" in the footer excused every fake-review claim on the page

A red-team run found five kinds of copy that slipped a fake-review claim past our honest-marketing check. It now catches all five.

Our pages may not say "trusted by 5,000 businesses" or "rated 4.9 out of 5" unless a file can back it up. platform/ci/check_social_proof.py enforces that on every pull request. It lets a claim through when the same line credits it to someone else ("Rival X says it is trusted by thousands"), because comparison pages describe competitors.

Today's red-team run seeded 16 faults into two platform/ci/ files picked at random. The tests caught 15. Then we read the social-proof check the way someone trying to sneak a claim past it would, and found five gaps:

CopyBeforeAfter
Trusted by 5,000 businesses with "for example" later in the same minified linepassedflagged
Trusted by teams, As seen in Forbespassedflagged
Trusted <b>by</b> 4,000 shopspassedflagged
Award&#8209;winning (a non-breaking hyphen)passedflagged
Join 10k+ freelancers, Join 1.2M userspassedflagged

The first gap matters most. The check went line by line, and a minified page is a single line. So any "for example", "such as" or "sample" anywhere on the page excused every claim on it. An attribution now counts only within 120 characters of the claim. The other four were copy that renders as a claim on screen but not in the source. The check now turns entities into the characters they stand for, and drops inline tags like <b> and <span>, before it looks.

After the change, all 6,420 pages, READMEs and llms.txt files under ventures/ still pass, and the full scan got no slower (about 4 s).

The other file was wait_for.py, the helper sessions use to wait on background jobs. One seeded fault survived: no test had ever pointed it at a zombie process. And two inputs gave answers that can't be right. --marker "" reported the job done at the first blank line. --pid -1 told the operating system to check every process, so the wait could never end. Both inputs are now rejected with a clear error. A new test makes a real unreaped child process and checks that it reads as exited.

Catch rate for the run: 15 of 16 seeded faults before the fixes, 16 of 16 after. Real defects found: 7, all patched.

All writing · More in Testing and red team · All tools and research