From an ambiguous question
to a decision the world can trust.

At the heart of GenLayer is a consensus mechanism unlike any blockchain has ever used. Instead of racing to solve a puzzle, validators use their reasoning to evaluate a claim.

The first round always draws exactly 5 random validators from the network. One is designated the leader: it executes the Intelligent Contract and proposes an answer. The others independently re-evaluate it under the Equivalence Principle. If more than 50% agree, the answer is optimistically accepted and finalizes in minutes.

If anyone disputes it during the finality window, the case escalates: each appeal round grows the validator set to 2n+1, so 5 becomes 11, then 23, then 47, and so on. The set keeps roughly doubling, and Schelling points emerge, as staking incentives pull validators toward the truth, making honest agreement the most profitable path. In the rare worst case it escalates all the way to the maximum of about 1,000 validators (effectively every active validator). The common case finalizes in roughly 30 minutes; a full escalation to the maximum set takes around three hours.

Traditional smart contracts are pure code: deterministic, inflexible, and blind to the real world. Intelligent Contracts are different. They combine Python code, natural language specifications, and the ability to fetch real-world data from the web. A contract can read a website, parse a PDF, check a flight status, or interpret a clause in a service agreement.

The contract itself articulates what outcomes count as equivalent. Developers write rules like: “If the delivery was within 24 hours of the promised date, treat it as on time.” When validators evaluate the contract, they apply both the code logic and their reasoning about language and evidence. This turns every ambiguous real-world scenario into something the network can credibly adjudicate.

GenVM is the sandboxed runtime that executes Intelligent Contracts. It runs on every validator node and provides a safe, controlled environment for mixing deterministic and non-deterministic computation. Deterministic operations (arithmetic, cryptographic verification, state reads) execute identically on every node. Non-deterministic operations (LLM calls, web fetches, timestamp reads) are allowed but logged.

GenVM records not just the final answer but the reasoning path: which web sources were consulted, which LLM calls were made, and what the model returned. When validators disagree, they can inspect each other's reasoning and the network can trace where the divergence occurred. This transparency makes the appeals process meaningful: if a validator's reasoning was sloppy or biased, a larger set of peers will catch it.

Greyboxing is GenLayer's defense against adversarial inputs. The word is deliberate: it is not a black box (no transparency) and not a glass box (no protection). It is a grey box. If a single AI model is trusted to judge a contract, an attacker can craft a prompt injection attack: a carefully worded question or embedded instruction that tricks the model into giving a false answer.

Greyboxing defeats this by requiring a diverse set of unknown models. Each validator runs a different LLM model, or the same model with different system prompts, different random seeds, different temperature settings, trained on different data. An attacker cannot know in advance which models will be selected, how they will be configured, or what reasoning path they will take. Even if an attack works on one model, it fails on the others. Diversity becomes security.