Contact

Defining Agent-Directed Manipulation · 日本語

Who publishes this

Perpensum is building third-party evaluation of AI agent purchasing decisions — the question of whether a machine's buying judgment was sound, as distinct from whether it was authorized. Audit logs and signed mandates answer the second question. They do not answer the first.

The definition published here is one piece of that: before you can say a page misled an agent, you have to be able to say what counts as misleading, in a way that anyone can check.

Perpensum does not accept payment, partnership, or preferential terms from any party in exchange for how that party is evaluated under this definition. That commitment is the product, not a courtesy.

How to reach us

GitHub issues are the channel: github.com/perpensum/agent-directed-manipulation/issues

Disagreements about the definition are especially wanted. If your implementation reaches a different verdict on the same input, that marks a place where the definition is ambiguous — which is a defect on our side, not yours.

There is no email address on this page yet. Rather than publish one we do not monitor, we say so. If you need to reach us and cannot use GitHub, open an issue is currently the only route — and that is a gap we intend to close.

If you want your own site watched

You can check any page yourself right now, with no involvement from us:

curl -s https://yoursite.com | node reference/scan.mjs

A one-off check is not the useful thing. The useful thing is being told when something changes — when a review section starts carrying instructions aimed at machine readers, or when hidden text appears on a page that was clean last month. Third-party regions are exactly where you are not looking.

Perpensum is building that continuous check. It does not exist as a product yet. There is no date, no price, and no waitlist that guarantees you anything.

If you want it, say so here: comment on the monitoring issue. No form, no email address, no signup — a GitHub comment is the whole mechanism.

This project's claim is that a judgement about someone else's site should be reproducible and honestly bounded. A vague promise to a prospective user would be the same failure in a different place. So: no product yet, and no dates.

If our crawler visited your site

It should not have. Perpensum currently crawls only domains its operator owns. If you saw the user agent below on a site you run and you are not us, something is wrong and we want to know.

Perpensum-crawler/0.1 (+https://perpensum.org/)

How it behaves, and what it will not do:

It obeys robots.txt. No exceptions, no separate rule for ourselves
It waits at least five seconds between requests to any one domain
It identifies itself. It never claims to be a browser
It stops on 403 or 429 for that domain, and does not retry around it
It does not evade bot detection and does not solve CAPTCHAs. If it is detected, it stops rather than working around it
It reads public pages only, and operates from an allowlist: we fetch what we have specifically approved, not everything we are not forbidden

To exclude it explicitly:

User-agent: Perpensum-crawler
Disallow: /

You should not have to ask. A crawler that only stops when challenged is not behaving well, and a company whose business is judging other people's conduct on the web has no standing to cut corners on its own.

Licensing

The definition and its documentation are CC BY 4.0; the conformance runner and reference implementation are MIT. Implementing the definition requires no permission, notification, or certification from us.