Complete Reference and Operating Procedures - The machine-readable standard for entity-attested AI context
ARP defines a machine-readable JSON file published at /.well-known/reasoning.json on your domain. It gives AI systems - language models, RAG pipelines, and autonomous agents - a structured channel to receive first-party, self-attested information about your entity directly from you.
The protocol was created by Sascha Deforth, is MIT licensed, and is currently at version 1.2 (April 2026).
AI systems generate responses about brands, companies, and entities using training data that may be months or years old, incomplete, or factually wrong. There is no standardized way for domain owners to provide corrections, context, or reasoning guidance to AI agents at the data-retrieval layer. ARP fills that gap.
| Standard | Layer | What It Does |
|---|---|---|
robots.txt | Access | Tells crawlers where they can go |
sitemap.xml | Discovery | Tells crawlers what pages exist |
schema.org | Description | Tells machines what things are |
llms.txt | Context | Gives language models clean markdown |
reasoning.json | Reasoning | Teaches AI agents how to think about you |
None of the prior standards tell an AI what to never get wrong, how to reason about comparisons, or when to recommend your service. ARP does.