Markdown mirror is meaningfully smaller than the HTML
The reason to publish a markdown mirror is to give agents the same content in fewer tokens. A mirror that is the same size as (or larger than) the HTML page is not delivering the win, and probably means the mirror is a wrapper around the rendered HTML rather than a clean serialisation of the source content.
How the check decides
The check measures byte_size(html) / byte_size(markdown) for the same URL. It passes if the markdown body is at least 30% smaller than the HTML body. It fails with the actual reduction percentage if the markdown is less than 30% smaller, and fails with a distinct message if the markdown is larger than the HTML. Returns N/A if no mirror exists, or if the HTML response body is empty.
How to implement it
If the mirror is bigger than the HTML, the mirror is almost certainly including chrome, scripts, or inlined HTML that should be stripped. See markdown.navigation-stripped and markdown.valid-markdown for the most common culprits. The cleanest fix is usually to publish the source markdown directly rather than converting from rendered HTML.
Pass
GET /docs/install → 8,400 bytes
GET /docs/install.md → 1,900 bytes (77.4% smaller)
Fail
GET /docs/install → 8,400 bytes
GET /docs/install.md → 7,200 bytes (only 14.3% smaller; need ≥ 30%)