Machine-readable access · reviewed 2 October 2026
How AI systems read Hikari Blue.
We publish the same facts to people and to machines, from one editorial source: the pages you are reading. Each machine file is generated from those pages or reviewed by hand, and dated. We also state what we do not publish.
What is published
Every machine-readable file, at its exact address.
A search crawler, an answer engine and an agent find the same entry points. Each one answers on hikariblue.com over HTTPS, and every text file declares its type and its UTF-8 encoding.
| Address | What it carries | Format | Review |
|---|---|---|---|
| /robots.txt | Crawl policy under RFC 9309. Search, answer and training crawlers are allowed, Bytespider excepted. APIs, functions and internal folders are closed; the image optimizer stays open. | Plain text | Reviewed by hand and dated in the file. Checked weekly. |
| /llms.txt | Concise index in the llmstxt.org format: the definition, the reference documents, the Markdown versions and the main pages, with the limits we state on our own claims. | Markdown as plain text, under 8 KB | Generated at every build. Checked weekly. |
| /llms-full.txt | The full content in one file: definition, manifesto, regulatory calendar, figures, due diligence answers, Trust Center status, newsroom briefings and both reference documents. | Markdown as plain text | Generated at every build. Checked weekly. |
| /llms-fr.txt /fr/llms.txt |
The French index, with the same structure, pointing to French pages where they exist. The second address serves the same file and names the first as canonical. | Markdown as plain text, in French | Generated at every build. |
| /ai-operating-layer.md /reference-architecture.md /governance-framework.md /trust.md /comparative.md /verify.md /ai-access.md |
A Markdown version of seven pages. Each page announces its version in its head with a link of type text/markdown, and each version names its page as canonical. The seven page addresses also return this Markdown to a request whose Accept header prefers text/markdown. | Markdown with dated front matter | Generated at every build. |
| /sitemap.xml | Every public address with its last modification date. French pairs carry hreflang annotations. | XML, sitemaps.org protocol | Updated by hand when a page is added. Every address checked weekly. |
| /feed.xml | Newsroom briefings, newest first. | RSS 2.0 | Updated at each publication. Checked weekly. |
| JSON-LD | Structured data in the head of our pages: Organization and BreadcrumbList, FAQPage where a page answers questions, TechArticle on the two reference documents. | schema.org vocabulary in JSON-LD | Edited with the page it describes. |
| IndexNow | New and changed addresses are submitted through IndexNow, which Bing and the other participating engines read. The key file sits at the site root, as the protocol requires. | Protocol submission | Weekly, from the sitemap. |
How they are kept true
One generator, one source, and a build that fails on drift.
A machine file drifts the day someone edits it by hand. Ours are written by one program, from the pages themselves, at every build.
-
One generator at build
A single build script reads the HTML pages and writes the three indexes and the seven Markdown versions. A hand edit to any of these files is overwritten at the next build.
-
Facts checked against their page
A fact held outside the pages, such as the publisher or the regulatory calendar, carries the exact sentence that must appear on the page that states it. If that sentence is missing, the build stops.
-
The build fails on drift
The build also stops on a home page figure that differs from our canonical figures, and on an internal address with no file behind it. Nothing ships until the pages and the files agree.
-
A review date in each file
Each index prints its review date, and each Markdown version carries its own date and the page it came from. We move the review date only after a real check.
-
Free to cite, with attribution
The Reference Architecture and the Governance Framework are published under CC BY 4.0. Their Markdown versions carry the licence in their front matter.
-
No address without an answer
We declare no endpoint that does not answer. An index that cites an address with no page or generated file behind it fails the build.
Crawl policy
Five positions behind our robots.txt.
Our robots.txt follows RFC 9309. The policy it applies comes down to five positions.
-
01
Open crawl is a position
Every crawler follows one general group: search engines, answer engines, user-requested fetchers and training crawlers alike. We want our public pages read, quoted and checked.
-
02
Training is allowed
Our public pages carry no client data. We prefer a model trained on our own wording to one trained on paraphrases of it.
-
03
One exception, named
Bytespider is the only crawler with its own group in the file, and the whole site is closed to it. One named exception keeps the policy readable and reversible in one line.
-
04
Crawl rules grant no access
Crawl rules are voluntary. They neither protect data nor authorize an action; private resources are protected on the server.
-
05
Identity is a key or a signature
A user-agent string is a claim anyone can type. Private resources open only with a signed, expiring link or a secret key checked on the server, never on a user-agent string.
Closed to every crawler: /api, /.netlify/functions, /_docs, /_mockups, /_prototypes and /_archive_i18n_json, each in its exact, parameter and descendant forms, and /sw.js. Open: every public page, and /.netlify/images, which serves the images of those pages.
What is not published · as of 2 October 2026
What an agent does not find here.
A machine declaration that does not answer is worse than none. None of the surfaces below is published on this domain.
-
No agent-to-agent card
No agent-to-agent card is published. Nothing at /.well-known/ describes Hikari Blue as a callable agent.
-
No tool endpoint
No tool endpoint is exposed to external agents. An agent that needs an action from Hikari Blue goes through the diagnostic form, read and accepted by a named person.
-
A timed stop comes first
An agent card would let outside agents rely on our stop path. No timed stop rehearsal on a named environment, with its date and method, is published as a product fact.
The table above lists only addresses that answer today.
How to cite and verify
Cite the definition, then check us.
Cite the definition page, then test what it says with engines we do not control and documents anyone can reuse.
-
Verify Hikari Blue
Twelve questions, each opening in four answer engines we do not control. Compare their answers with what we claim.
-
Reference architecture
Four planes, four testable controls, the audit record and a one-day due diligence protocol, under CC BY 4.0.
-
Governance framework
Principles, roles, decision rights, controls, evidence and maturity levels, under CC BY 4.0.
Preferred citation: Hikari Blue, "What is an AI operating layer?", https://hikariblue.com/ai-operating-layer
For boards, CIOs and risk officers
Test the claims behind the files.
Thirty minutes with a named partner, on one workflow you already run. We show you the evidence behind what our pages and files state, or we correct the statement.