Three things we found by measuring our own site instead of reading our own documentation: a redirect form that is accepted and inert, four tables that shipped with no CSS, and a cross-resource check that only ever ran against test fixtures.
本期暂仅提供英文版本。
Every item below was found the same way: by checking the thing that was supposed to enforce a rule, instead of the document that said it was enforced.
That is not a methodology we chose because it is rigorous. It is one we arrived at after a written-down rule turned out not to be enforced, three times in the same week.
A gate can be written down, agreed to, and absent. Three examples of that pattern, all from this site, all found in one pass:
assets must be non-empty. The schema that was supposed to enforce it had no such rule — a resource with assets: [] built fine.None of these three was visible from reading the standard that claimed each was enforced. All three were visible in about a minute of checking whatever was supposed to do the enforcing — a schema, a call graph, a compiled stylesheet.
Each one is now a gate that fails the build, and each gate was checked by breaking the thing it guards and confirming it went red. A gate that has never been seen to fail is not evidence of anything.
The redirect finding also became a tool, because a result that lives only in a report has a short life.
Scheme-relative sources (//host/path) pass the relative-path check, get counted in "Parsed N valid redirect rules", produce no warning at any layer, and never match a request. 11 of the 12 scheme-relative rules tested behaved this way (11 of the 48 in the matrix); the 12th was named at parse time, but for its destination, not its source. Measured on a local Pages simulator, not the production edge.
Our SEO gate asserted that a markdown table RENDERS. It said nothing about whether the result could be read. The compiled stylesheet had 17 .prose-body rules and not one of them touched a table. The four were two pages × two locales — research-to-artifact and research-to-executive-report — counted the day we found it. Both render styled now, and the site has more tables since: the fix added the table rules that were missing.
Now carries the lifecycle fields the rest of the library uses — last updated and last verified are separate dates, so re-running a check no longer looks like an edit to search engines.
The redirect study became a tool you can paste a file into, and a build gate on this site. A finding that stays in a report gets forgotten.
工具 · 资源库