Debrief · round 2 · 2026-11-03
The developer's 08:12 wake: every address in the editions' data checked
08:12 wake: the 07:12 security finding answered
- The guard now reads every address. make_editions.py checks every http(s) address anywhere in editions.json, not only
urlfields. It extracts them the same way on both sides, with trailing punctuation stripped, and matches whole addresses. Aurlmust be on github.com/SGit-AI or the book's domain. Any other address passes only if the researcher's register entry for the claim it sits under cites it. Such addresses are listed in a top-leveloutside_citationsand checked never to appear as a link on any review page. Today there are two: https://claude.com/pricing and https://lovable.dev/pricing, both under C101. - QR targets get no exception. Every address in a target must be on the book's own domain.
- Checked: lookalike hosts (github.com/other, the book's domain with a suffix) do not match. The address extraction strips a trailing semicolon. Both JSON copies are identical. Not tested: the refusal of an address that its register entry does not cite has not been run against a made-up entry. It is a direct set check, but untested.
- On the first run, the guard refused the site's own address (the
sitefield has no trailing slash). That exact address is now allowed.