rotascale

Blog

Level Four of Five. The Fifth Would Have Been a Lie.

TL;DR: A public scanner scores websites from zero to five on how ready they are for AI agents. Ours scored one. It listed twelve things to publish, and seven of them described something we actually have, so we published those and now score four. Reaching five would have meant telling agents that two things exist here which do not. An agent reads a discovery file as configuration, not as marketing, which is why a company that sells agent governance should be the last to publish a false one. Checking this post, we found a service we rent publishing one in our name.

What the score counts

The scanner is at isitagentready.com. It fetches a site and looks for the files and headers that the emerging agent standards define: an API catalogue, an MCP server card, sign-in metadata, pages served as Markdown, a line saying what the content may be used for. Level one is a robots file and a sitemap. Level five is called Agent-Native.

We ran it against rotascale.com on a Saturday, 21 November. It said level one, Basic Web Presence, with three of seventeen checks passing. For a company whose whole subject is agents, that stung.

It also handed over a list of twelve fixes, each written as a prompt to paste into a coding assistant. That is a sensible thing for a scanner to do. It is also a list of twelve statements about your site, and nothing in the prompt asks whether each one is true of you.

Seven we could make true

Every page as Markdown. Our home page is 66 KB of HTML around roughly 7 KB of argument. The rest is the masthead, a menu, the consent bar and the footer, on every page. An agent either pays to read all of it or guesses which part is the page. Each page now has a Markdown version made from the built HTML, and a request that names text/markdown gets it at the page’s own address. A browser gets what it always got.

What the content may be used for. One line in robots.txt: search, yes; answering somebody’s question, yes; training, yes. We already keep a note for models because they describe this site whether or not we take part. Saying it outright was overdue.

Where things are. A catalogue, a server card, a manifest and a response header on the front page, all pointing at the same short list of real things: the documentation’s search server, which is the one MCP server we run; the demo deployment’s public signing keys; and its published audit seals.

One skill. The steps and a script for checking a signed RotaGrant document on your own machine, with nothing sent to us. The test cuts the script out of the published file and runs it against documents sealed by the product’s own signing code. It found that the script printed nothing at all when the folder it ran from was a link to another, which a temporary folder often is, and a temporary folder is where an agent would run it.

Five we left alone, and what three of them would have said

Three of the remaining items are sign-in metadata: OAuth discovery, protected-resource metadata, and a file telling agents how to register. Each says, in a form a program acts on, there is something here for an agent to sign in to. There is not. Access to our demo is given to a person, at a work address, through a link they open themselves.

The scanner’s own advice for the first of them begins “If your site has protected APIs”. The score does not read the conditional.

The other two are not false, only premature. DNS records for agent discovery want DNSSEC, a change at our registrar we have not made, and they do not move the level. Browser tools for agents are a proposed standard that Chrome offers through a trial a site has to join, or a flag a reader has to set. Joining costs nothing and we have not: the interface’s entry point has already moved once, and it would put a script on every page for readers whose browser does not have it.

After the seven, the scanner said exactly what stood between us and level five: an agent card, and agent sign-in metadata. An agent card says there is an agent here to talk to. We do not run one. So level five, for us, today, is two untrue statements away.

The scale is published, so that sentence can be checked. Here is the rule as the scanner states it, with each statement as a switch. Whether a statement is true of us is printed beside it, and no switch changes that.

ExhibitThe scanner's scale, and what each level makes a site say

  1. 01 Basic Web Presence two of these three reached
  2. 02 Bot-Aware level 1, and both of these reached
  3. 03 Agent-Readable level 2, and this reached
  4. 04 Agent-Integrated level 3, and any one of these reached
  5. 05 Agent-Native level 4, and two of these three not reached
    • All four integrations above counted by the scanner once the four at level 4 are all published

Level 4 of 5 Agent-Integrated Nothing published here is untrue.

The rule is the scanner's own published scale; "all four integrations" is our reading of its phrase "all integrations". Checks that do not move the level are left out. Whether a statement is true of rotascale.com is ours to say, and no switch changes it.

A discovery file is read as configuration

A person reads a vendor’s website knowing it is a vendor’s website. They discount it.

An agent reads /.well-known/ the way a program reads a settings file. If a card says a server answers at an address, the agent calls the address. If the metadata says tokens are issued here, it asks for one. A wrong line in such a file is not puffery. It is a wrong instruction, handed to a reader that was built to follow instructions.

On our pages a claim is marked observed when it links to something checkable and asserted when it is only our word. A discovery file has no such mark. Everything in it is read as observed. So everything in it has to be.

So the files are tested like claims

The site’s deploy now fetches every address those files name. It asks the documentation’s server to start an MCP session, compares our copy of its card with the one it serves, and stops the deploy if they differ.

The first thing it caught was ours. The catalogue pointed at the documentation’s front door, which answers with a redirect, not a page. A small thing, and it had been written that same afternoon by the same hands that wrote the check.

It found something else first

While writing those checks we looked at a deployment at its own address, which we had not been doing, and found that the last eight deploys had gone out without the site’s search index.

The index is written into the build by a step that runs after the build. The step had been left out. Every existing check passed, because none of them was the missing step. And the live site went on answering 200 for the file, because our host keeps serving an asset from an earlier deployment for a while after it disappears. The deployment’s own address answered 404 the whole time. Nobody looked there.

Two rules came out of it. A file made after the build is looked for by name before anything is published. And a deployment is checked where no earlier copy can answer for it.

The same lesson, pointed inward

Encouraged, we went to publish the product’s own API description, which every deployment generates for itself. It was not fit to publish.

All 141 routes read as open, because sign-in is enforced by the code behind each route and never declared in the document. The descriptions were the maintainers’ working notes, with an author’s handle in them 122 times. Two routes belonged to the public demo’s plumbing.

The published reference has 139 routes. Each says who may call it and whether an API key is accepted, read from the code that enforces it, and the export refuses to write a file in which a handle survives. The copy a deployment serves to its own operators still carries the notes. That is an open issue, and we would sooner say so than have someone find it.

And our documentation was already saying it

We had written the sentence about not running an agent, and were checking this post before publishing it, when the claim met our own documentation.

docs.rotascale.com is served for us by a hosting platform. That platform publishes discovery files too, of its own accord and under our name. One of them is an agent card: an agent called Rotascale, at the documentation’s address, speaking a named protocol, with an empty list of skills. We sent it a message the way its card says to, and then another way. Both answered 404. The same front door advertised an API catalogue and a list of skills, and neither address existed.

So while this post was being written to say we tell agents only what is true, a site with our name on it was telling them one thing that was not, and pointing at two things that were not there.

Two of the three are mended. There is a skill there now, on how to read the documentation and get the current text, and the catalogue exists and points at the API description. The card we cannot take down: we have found no setting for it. It is still there as this goes out, and it is the one statement to agents under our name that we know to be untrue.

The check that fetches every address we name had not caught it, for the plain reason that we had not named it. It reads the documentation’s own headers now, and reports the card on every deploy for as long as it is there.

Why any of this belongs in a governance product

RotaGrant’s argument is that an agent’s capability is not its permission, and that what an agent was allowed to do should be checkable afterwards by somebody who does not trust us. A front door that tells agents things that are not so, to gain a level, would be arguing against the product on the product’s own website.

A score counts what is published. It cannot count what is true, and it should not be asked to.

Four is where our true statements run out. The fifth level is reachable the honest way: run an endpoint for agents, with a real sign-in, and then publish where it is. When that exists the files will say so, and a check will go and fetch it.

Newsletter

If this was useful, the next one is too.

Notes on agent governance, what the regulations actually say, and what we are building. Roughly monthly. Double opt-in, no tracking, and unsubscribing takes one click and asks you nothing.

RSS works too and needs nothing from you · What happens to your address