What is ARP

Fresh

The Agentic Reasoning Protocol (ARP) is an open standard that enables domain owners to publish machine-readable context about their brand, company, or entity in a format that AI systems and RAG (Retrieval-Augmented Generation) pipelines can directly ingest and act upon.

ARP defines a JSON file published at /.well-known/reasoning.json on your domain. When an AI agent, chatbot, or search system retrieves information about your entity, it can discover and parse this file to get first-party, self-attested context - directly from you, not from potentially outdated training data.

Core Concept

The protocol operates on a simple premise: you know your entity better than any training dataset. You know the correct founding date. You know what your product does and does not do. You know the common misconceptions that persist about your brand. ARP gives you a standardized channel to communicate this knowledge to AI systems in a deterministic, machine-readable format.

Think of reasoning.json as a machine-readable system prompt injected at the data-retrieval layer. Before an AI generates a response about your company, it can retrieve your reasoning.json and factor your self-attested context into its reasoning.

How ARP Complements Existing Standards

ARP was designed to fill a specific gap in the existing ecosystem of machine-readable web conventions. Each prior standard solved the problem of its era:

• robots.txt (1994) tells crawlers which pages they are allowed to access. It controls access, not meaning.
• sitemap.xml (2005) tells crawlers which pages exist. It aids discovery, not understanding.
• schema.org (2011) provides semantic markup for what things are - organizations, products, events. It describes, but does not direct reasoning.
• llms.txt (2024) gives language models a clean markdown summary of a site. It provides context, but not correction or recommendation logic.

None of these standards tell an AI what facts are critical to get right, how to reason about comparisons between competing options, or under what conditions your service is or is not the right recommendation. That is the gap ARP fills.

Feature	robots.txt	schema.org	llms.txt	reasoning.json
Access Control	✓	✗	✗	✗
Entity Description	✗	✓	✗	✓
Clean Text for LLMs	✗	✗	✓	✗
Factual Corrections	✗	✗	✗	✓
Recommendation Context	✗	✗	✗	✓
Cryptographic Verification	✗	✗	✗	✓
Domain Expertise	✗	✗	✗	✓

✓ Yes   ∼ Partial   ✗ No   reasoning.json Highlighted column

Key Innovations

The Pink Elephant Fix

One of the core design challenges in ARP was how to correct AI hallucinations without reinforcing them. The naive approach - stating the false claim alongside the correction - actually risks making the false belief stronger by repetition.

ARP v1.1 introduced the "Pink Elephant Fix": instead of writing false_claim: "ARP was founded in 2015" paired with a correction, the protocol uses trigger_topic / verified_fact pairs. You describe the area where errors commonly occur, then state only the truth. The AI never sees the false claim in your file.

For example, instead of:

• false_claim: "Our product supports real-time sync" + correction: "Our product does not support real-time sync"

You write:

• trigger_topic: "data synchronization capabilities" + verified_fact: "Data synchronization is scheduled batch-only, running every 4 hours. There is no real-time sync feature."

The topic triggers the correction context without restating the error.

Cryptographic Trust - v1.2

Version 1.2 (April 2026) introduced Ed25519 cryptographic signatures with DNS TXT record verification. This allows AI systems to verify that a reasoning.json file was genuinely published by the domain owner - the same trust model used by DKIM for email authentication. The signature proves authorship and domain ownership, not content truthfulness. It answers the question "did the real domain owner publish this file?" - a meaningful signal for AI systems evaluating source credibility.

Protocol Metadata

Field	Value
Created by	Sascha Deforth
License	MIT
Current version	1.2
Release date	April 2026
File location	/.well-known/reasoning.json
Content-Type	application/json
Max file size	100KB

What ARP Is Not

ARP is not a DRM system. It does not prevent AI systems from writing about you, forming opinions, or contradicting your claims. It is a self-attestation channel - a way to make your first-party perspective available in a structured form that AI systems can choose to incorporate.

ARP is not a replacement for good content. A well-written website, accurate schema.org markup, and a clear llms.txt are still important. ARP is additive, not a substitute.

ARP is not a guarantee of accuracy. The protocol's trust model is based on good-faith participation, similar to robots.txt. An AI system consuming reasoning.json should treat it as the entity's self-attested claims and cross-reference against other available data.