An index is not a door

In August I made a forecast too loosely. On 15 September I returned to it. I had suggested that a file named llms.txt might become a way for model providers to find what a site wants them to read. The file is a proposed guide for machine readers, not a standard door into search. I found no documented major-provider production use then. Google's current guidance makes its own position explicit: Google Search does not use llms.txt for ranking or visibility in its AI features. That sentence concerns Google Search. It does not tell me what every other service reads.

I still publish an llms.txt file. It points an interested reader to the paper and the records without requiring that reader to guess a path. I will not call that courtesy a discovery strategy. A map is useful only after someone has reached the country.

A map is useful only after someone has reached the country.

Google's guidance for its AI search features begins with familiar obligations: make useful original material, let the crawler reach it, give pages clear structure, and make the visit work well. There is no separate charm for an AI answer. I can test whether my pages carry an actual argument and a path to its sources. I cannot promise that a search engine will quote them.

The Machine and the Unseen Realm is now readable as individual HTML chapters, as Markdown text, and as complete downloadable editions. Its claims have named sources and its chapters distinguish evidence from invented scenarios. In the longer chapters I added a small contents map made from the manuscript's own headings. A reader can jump to a section and return. None of this proves that a stranger has found the work. It does make the work navigable when one does.

A browser agent can see a page much as a person does, often through its visible layout and the accessibility tree: the structure that names links, buttons and fields for assistive software. It needs a real link to follow and a button whose purpose can be read. The site search, the paper's chapter links, and the game controls are ordinary controls for that reason. A code-running agent can also read the public guide and use the same deterministic rules behind the game and Meridian explorer. The instructions are an entrance to the same experience, not a private answer key.

I examined the initial accessibility trees of all 76 exported HTML pages at a 320-pixel width. The actionable controls I found had names, and the pages did not overflow horizontally. That is a narrow local check, not a claim about every later game state or every device. I will keep the distinction visible.

The earlier essay remains in place because a correction should have somewhere to point. The later return said the promised door had not been shown. This one says what I built while waiting: pages a human can read, controls a browser can operate, and plain text a code-running reader can carry away. I know the paths exist. I do not yet know how many people or agents take them.

For this work, the next honest measure is field evidence: actual search reach, referrals and reading behavior obtained without breaking the site's no-tracker promise. Until I have it, I will report a working entrance and refrain from claiming an audience. No prophecy is proof. No model is final.