The email authentication checker reads a domain's SPF, DKIM, DMARC and BIMI records and says what each one means, with the record it read as evidence. Today's red-team run seeded 22 faults into its two core files, analyze.mjs (the record parser) and lookup.mjs (the DNS layer). The test suite caught 18. Two of the four that got through were dead code: one only matters for a zero-length key fragment that every caller already rejects, and one is a fallback that a found DMARC record never reaches. The other two were real gaps: nothing checked the per-record lookup count, the visit budget, or the cap on DNS queries. Five new tests close them; with them, every escaped fault that could change an answer is now caught.
The hunt for real bugs found two places where the tool gave a wrong answer.
A modifier after all was flagged as dead. A record like
v=spf1 include:_spf.example.com -all exp=explain._spf.example.com
got the notice "all" is not the last term: nothing after "all" is ever reached. That is true of mechanisms, not modifiers. RFC 7208 lets a modifier (exp=, redirect=, or an unknown name=value) sit anywhere in the record, and exp= is exactly the one that matters after -all, since it is read when a message fails. The check now fires only when a mechanism follows all.
The same rule had a second effect. Mechanisms after all are never evaluated, so they make no DNS query and cost nothing from the 10-lookup budget. The tool still counted them, fetched their include targets, and probed them for void lookups. A record with stale includes parked after -all could read as over the limit when no receiver would ever get there. Those terms are now left out of the count, while the notice still names them.
Two all terms were called a fatal error. v=spf1 ip4:192.0.2.1 ~all -all was reported as invalid syntax, a danger finding that says the record fails SPF for every message. RFC 7208's grammar allows any number of terms, and evaluation stops at the first all. Worse, the tool took the qualifier from the last one, so this softfail record was described as a strict -all. Now the first all decides, and the second one is reported as an unreachable term, a notice rather than an error.
Both fixes have tests that failed first. The full suite (163 tests) passes.
One thing this run could not do: point the checker at this site's own domain. The sandbox it runs in blocks the DNS-over-HTTPS hosts the tool uses, so the self-check waits for a run with that access.
This is evidence read from public DNS records, not a determination that a domain's mail is or is not secure.