---

README for the RFC-AI Initiative Repository

---

Welcome to the RFC-AI Initiative Repository!

This repository serves as the official home for the drafts (AI-Drafts) and published documents (RFC-AIs) of the RFC-AI Standardization Initiative. Our mission is to foster a more interoperable, secure, accountable, and transparent Artificial Intelligence (AI) ecosystem. To achieve this, we develop open, consensus-driven technical standards for AI, Large Language Models (LLMs), multi-agent systems, and related technologies.

Inspired by the successful Internet Request for Comments (RFC) model, our goal is to provide clear, implementable specifications that complement existing standardization efforts. All our documents are managed in Markdown and version-controlled using GitHub, which facilitates open collaboration and agile development.

We invite you to explore our documents, contribute your ideas, and actively participate in building a more standardized future for AI.

---

Document Index (RFC-AIs)

Here you'll find a list of RFC-AI documents, categorized for easier navigation. The RFC-AI numbers listed are the definitive identifiers for each document.

Terminology and Definitions Standards (Informational RFC-AIs)

These documents aim to establish a common vocabulary and clear definitions for fundamental concepts in the field of AI.

• RFC-AI 0000: Statement of Purpose and Introduction to the RFC-AI Standardization Initiative

  • This foundational document explains the motivation, vision, and scope of the RFC-AI Initiative. It's the ideal starting point for understanding our work.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0001: Fundamental Glossary of AI/ML/LLM/Agents

  • Clear, consensual definitions of key terms such as "Large Language Model (LLM)," "Autonomous Agent," "Multi-Agent System (MCP)," "Inference," "Training," "Hallucination (in LLMs)," "Bias (in AI models)," "Explainability (XAI)," "Prompt Engineering," etc. This glossary is crucial for precise communication and for building other standards.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0002: Taxonomy of Agents and Multi-Agent Systems

  • Classification and description of different types of agents (reactive, model-based, goal-based, etc.) and multi-agent system architectures.
  • Link to document (Hypothetical path in the repository)

Interoperability Standards

This category addresses technical specifications that enable different AI components, models, and systems to communicate and interact seamlessly.

• RFC-AI 0003: Common Formats for Model Exchange (Metadata)

  • Definition of standard metadata to describe AI models (type, architecture, size, training data, known performance metrics, hardware/software dependencies). This focuses on how to describe the model, not necessarily its binary format.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0004: Basic Agent Communication Interfaces

  • Definition of protocols or APIs for different agents or multi-agent systems to discover each other, negotiate tasks, and communicate information or actions in a structured way.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0005: Formats for Representing Agent Tasks and Goals

  • Standards on how tasks or goals assigned to an agent can be described in a way that is understandable to agents developed by different entities.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0006: Standard APIs for LLM Invocation

  • Specifications for common Application Programming Interfaces (APIs) to send prompts, receive responses, manage context, and handle sessions with LLMs, regardless of the underlying provider (if possible or applicable at a basic level).
  • Link to document (Hypothetical path in the repository)

Security and Robustness Standards

These documents focus on technical guidelines to enhance the security of AI systems and their resilience to attacks or unexpected behaviors.

• RFC-AI 0007: Security Considerations for AI Model Deployment

  • A document of best practices and minimum technical requirements to protect models against attacks (data poisoning, adversarial attacks, model theft) and to secure the systems where they run.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0008: Definition and Classification of Vulnerabilities in LLMs and Agents

  • A catalog of common types of vulnerabilities (e.g., prompt injection, data exfiltration, jailbreaking) and a framework for describing them.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0009: Basic Robustness Metrics Against Adversarial Examples

  • Standardization of metrics to evaluate how resistant a model is to small perturbations in input data designed to trick it.
  • Link to document (Hypothetical path in the repository)

Accountability and Ethics Standards (Technical Aspects)

This section is dedicated to translating ethical and accountability principles into implementable technical specifications.

• RFC-AI 0010: Technical Guide for Training Data Documentation

  • A standard on what information should accompany a dataset used to train models (origin, curation process, demographic/distribution characteristics, known biases).
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0011: Technical Metrics for Measuring Bias and Fairness

  • Definition of quantifiable metrics to evaluate different types of biases in a model's predictions or decisions, applicable in various contexts.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0012: Basic Formats for Reporting Model Uncertainty or Confidence

  • Establishes how a model can communicate, along with its prediction or response, a standardized indicator of how confident it is.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0013: Minimum Transparency Requirements in Agent Systems

  • Specifying what kind of "logs" or "justifications" an agent should be able to generate to explain its sequence of actions or decisions in response to a stimulus or task.
  • Link to document (Hypothetical path in the repository)

Evaluation and Benchmarking Standards

These documents focus on defining methodologies and formats for the objective evaluation and comparison of AI systems.

• RFC-AI 0014: Standardized Methodologies for LLM Evaluation (NLU/NLG Tasks)

  • Definition of protocols for executing and reporting results on common Natural Language Understanding (NLU) and Natural Language Generation (NLG) tasks, to make benchmarks more comparable and reproducible.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0015: Standardized Metrics for Agent Behavior Evaluation

  • How to measure and report the performance, efficiency, and reliability of agents in simulated or real environments, enabling meaningful comparisons.
  • Link to document (Hypothetical path in the repository)

• RFC-AI 0016: Formats for Exchanging Benchmarks and Results

  • Defines a standard structure for sharing evaluation datasets and the results obtained by different models or agents on them, facilitating transparency and reproducibility.
  • Link to document (Hypothetical path in the repository)

---

How to Contribute

Your participation is crucial to the success of this initiative! If you wish to contribute to the RFC-AI Initiative, we encourage you to:

1. Explore the Drafts: Review the AI-Drafts in the drafts/ folder to familiarize yourself with the documents currently under development.
2. Submit Issues: Open Issues on GitHub to propose new ideas, point out errors, suggest improvements to existing documents, or initiate discussions on relevant topics.
3. Submit Pull Requests: If you have direct contributions to the content of the drafts, submit Pull Requests. Please ensure you follow our style and formatting guide (to be detailed in a future RFC-AI).
4. Join Working Groups: Participate in relevant Working Groups (WGs) to engage in deeper technical discussions and help shape the standards. Consult our document on the Initiative's process for more details.

---