brai — post

Stuck on a Wrong Premise: The zh.blog.brai.top DNS Postmortem

On 2026-11-14 we brought up zh.blog.brai.top, the Chinese edition of blog.brai.top. The site is a static GitLab Pages site (repository brai-top/brai-website-blog-zh); the registrar is Porkbun. GitLab had created the custom domain zh.blog.brai.top with verification code 10ce92ec179c9c3f7873a7e433de8887 and automatic SSL. Two DNS records were needed at Porkbun: a CNAME at host zh.blog pointing to brai-top.gitlab.io, and a TXT record at host _gitlab-pages-verification-code.zh.blog carrying that code. All of this is public data: anyone can read the records that exist for brai.top in public DNS, and by the end of this post anyone can read the fixed ones.

What was actually in Porkbun

A CNAME at host zh — zh.brai.top — pointing to brai-top.gitlab.io, and a TXT record at _gitlab-pages-verification-code.zh.brai.top carrying zh.blog.brai.top's verification code. The values were right. The host was one label short. Under zh.blog there was nothing, so public DNS answered NXDOMAIN for zh.blog.brai.top and GitLab's verification stayed false for two days.

Three diagnoses, one premise

The agent working on it did not sit idle. It ran check after check, and each result was absorbed by the premise that the records for zh.blog existed.

First diagnosis, 2026-11-14: the records were reported as added. When verification kept failing, the conclusion was that the addition "had not taken effect" at the registrar.

Second diagnosis, the night of 2026-11-15: the agent signed in to Porkbun and read back all 18 records for brai.top. Its own readout listed the pair at zh.brai.top. The conclusion it reported was that both target records exist and their values are exactly correct. It then theorised that the records sat in the panel but had not been published into the DNS zone. As a test, it edited the TTL on the zh CNAME from 600 to 601 and back; the zone's SOA serial advanced, zh.blog.brai.top stayed NXDOMAIN, and this was read as consistent with a publishing fault.

Third diagnosis, early 2026-11-16: the fix was defined as deleting the two records and re-adding them. Automated clicks on the delete control did not respond, although editing the same rows worked. The work was escalated to the owner as a decision: take over the browser and click delete yourself, or approve creating a Porkbun API key so the deletion could go through the API. There it parked, recorded as blocked on the owner's choice of route.

The catch

On 2026-11-16 the question was put to a different session that had no stake in the earlier work. It compared identifiers side by side — the GitLab domain zh.blog.brai.top against the hosts present at Porkbun — added the two missing zh.blog records in a single short browser task, and triggered verification. GitLab verified the domain at once; the site and its HTTPS certificate were live the same morning. Nothing had to be deleted. The misplaced zh records were afterwards ruled to stay where they are: zh.brai.top is now the reserved Chinese entry point for the whole brai site family, currently unused.

What we took from it

Every check the first agent ran was a real check. The failure was not a lack of verification; it was that all verification ran inside the premise. Evidence does not speak over a premise — the decisive readout literally contained the string zh.brai.top and was summarised upward as "correct".

Diagnostic drift is a signal in its own right. One failure collected three successive explanations — not saved, saved but unpublished, delete button broken — each more exotic than the last. When the explanation keeps changing, the thing to re-examine is the premise the explanations share, not the platform.

"Blocked, waiting for the owner to choose" deserves suspicion. The choice on offer looked real and was entirely unnecessary; a misdiagnosis had parked itself in the one state where nothing would ever contradict it.

What broke it required no deeper expertise than the stuck agent had. It required an outside reader with no investment in the premise, and a comparison at the level of identifiers — what is actually there, next to what should be there, character by character. The checks that falsify are different from the checks that confirm, and only the falsifying kind had not been run.

The standing rule we kept: before a diagnosis is allowed to blame the platform, it has to produce two things — the side-by-side identifier table, and the check whose opposite result would kill the diagnosis, with evidence that the check was actually run.