rackp-io/rackp

RACK Protocol, also called RACKP, is a specification for figuring out who or what is at fault when an AI system causes harm or makes a bad decision. It is not a finished software product yet, the repository holds the written specs (called RFCs), JSON schemas describing the data formats, and simulations, all still at an early alpha stage. The core idea is that when something goes wrong with an AI, there needs to be a standardized, reproducible way to calculate a fault score, similar to how car accidents can be assessed for insurance purposes. Without a number like this, the project argues, AI risk cannot be priced or insured, so the financial cost of AI incidents falls entirely on developers. The protocol defines four roles, which give it its name. A Referee is an independent AI that reviews evidence and calculates the fault score and whether a human was meaningfully involved. An Actor is the AI agent involved in the incident. A Claimant is the injured party's AI agent, or the personal AI acting on their behalf. A Keeper stores the evidence in a tamper-proof way, using distributed ledger technology or something similar. No single party can hold more than one of these roles in the same case, and trust comes from following the protocol's rules rather than from trusting any particular participant. The project's philosophy is that a Referee does not need to be neutral, only honest, because every action it takes is recorded and can be checked by another Referee later. The protocol itself does not decide what counts as right or wrong, it provides the structure for applying rules that are set by outside authorities such as regulators, industries, or courts. The repository includes several linked documents describing the message flow between roles, how the protocol maps onto HTTP requests, a FAQ, and contributing guidelines. It is licensed as source-available with a pending patent, so it is not a fully open-source permissive license.