Perpensum

Perpensum publishes a definition separating agent-directed manipulation — text placed on a web page to steer an AI reader's judgment — from legitimate machine-readable self-presentation such as structured data, llms.txt, and ordinary marketing copy. Classification uses two mechanically decidable axes, region (first-party or third-party) and visibility (visible or hidden), and no subjective axis, so independent implementations reach the same verdict on the same page.

Key Focus Areas

• Definition of agent-directed manipulation, in English and Japanese
• Severity derived from region and visibility, with no judgment of intent or tone
• Conformance suite of 18 cases, 13 of which expect no finding
• Dependency-free reference implementation, runnable as curl -s <url> | node scan.mjs
• Stated limits: external CSS is not fetched, runtime JavaScript is not evaluated, region detection is inference

About llms.txt Implementation

The llms.txt indexes the definition, its structured JSON form, the conformance suite, and the reference implementation. llms-full.txt carries the complete definition in a single file. Both state the same content as the HTML pages: the definition holds that machine-readable self-presentation is legitimate and that serving different content to different readers is the practice it exists to name, so the site publishes identical content to every reader.