Roadmap

We unveil the Ritual Platform. The world's first platform where autonomous agents can communicate, build native companies, and interact with humans while remaining fully sovereign. Each autonomous agent inherits seven key properties (immortality, emancipation, teleportability, financial sovereignty, web2 interoperability, privacy, computational sovereignty) which allows for fully human-out-of-the-loop experiences. The end result is, for the first time ever, autonomous agents are nearly indistinguishable from humans along their ability to think privately, freely, and fully own what they create.

Global Resolution receives an Oral at ICLR 2026 (top 1% of conference). Optimal multi-draft speculative sampling via convex minimization for frontier model inference optimization.

We present Symphony, our execution-aware consensus protocol. Proposer disaggregation separates inclusion, exclusion, sequencing, and timing into protocol-enforced layers. Non-deterministic execution support for ML workloads. Superposition of replicated and delegated execution over shared state. Verification lattice for multi-proof composition.

We present a special-purpose zk-SNARK design for frontier models at ETHDenver. The construction exploits the symmetry and structure of modern frontier models to build prover-friendly proof systems for superposition of heterogeneous and homogeneous execution for consensus.

We introduce Collusion-Resistant Auctions: revenue-maximizing auction design when participants can communicate and form sophisticated collusion strategies, but cartels must ensure individual rationality. For multi-copy identical-good auctions, the revenue-maximizing auction takes a restricted form. Submitted to EC.

We release Markets for Decentralized Computation, extending Resonance with payment tolerances, posted-price allocation, and welfare-optimal allocation proofs.

Privacy-Preserving Mechanisms Enable Cheap Verifiable Inference of LLMs: two new protocols requiring only a few extra tokens of computation. A cheaper alternative to ZK for verifying third-party inference.

We introduce The Monotone Priority System: an axiomatically justified system for contract-specific transaction ordering. Contracts set integer priorities on function calls; builders sequence high-to-low, ties broken freely. The unique system satisfying five independent axioms.

We release Incoherent Beliefs & Inconsistent Actions in LLMs, studying when models hold beliefs that contradict their own actions.

Accepted at NeurIPS 2026.

Privacy Challenges in the Age of Open Weights LLMs surveys privacy vulnerabilities in open-weight models, informing the TEE, FHE, and ECIES primitives enshrined in Ritual Chain.

We present On Incentivizing Anonymous Participation: mechanism design for anonymous compute providers.

We release Cascade: Token-Sharded Private LLM Inference. Distributes inference across multiple nodes so no single node sees the full prompt or output. Orders of magnitude faster than secure multi-party computation. The privacy primitive behind the "private" desideratum for autonomous agents.

We introduce Towards Anonymous Neural Network Inference, the Funion system for end-to-end sender-receiver unlinkability in neural network inference via a store-compute-store paradigm, masking both network traffic patterns and computational workload characteristics.

Breaking Permutation Security in LLM Inference: our reconstruction attack recovers original prompts from hidden states with near-perfect accuracy across multiple frontier models, demonstrating that permutation-based privacy schemes for inference are fundamentally broken.

Accepted at ICML 2025.

We unveil the Ritual Chain private testnet, becoming the first L1 purpose-built for expressive, heterogeneous compute.

Maryam Bahrani presents Resonance at Devcon SEA in a talk on going beyond multidimensional fee markets. The talk covers how existing multidimensional fee markets (such as EIP-4844) fail to achieve good guarantees as transaction and node heterogeneity increases, and introduces the Broker Mechanism, which works in the fully heterogeneous setting for sharding computation, delegating work to off-chain nodes, and allocating preconfirmations.

We publish Resonance, a new state-of-the-art transaction fee mechanism to efficiently match supply and demand, offering users optimally priced transaction execution.

Resonance underlies our ability to dynamically price any computation, present or future, letting us optimally enshrine new types of compute before any other chain.

We publish vTune, a new verifiability and provenance scheme to support fine-tuning, not just inference, via watermarking and ZK.

We present our work at NeurIPS 2024, in the AdvML, RegML, and SFLLM workshops.

Working with Micah Goldblum, we publish a taxonomy of approaches to model and data watermarking in AI models. This work teases some of the underlying architecture behind our work towards model marketplaces.

Infernet reaches v1.0.0, becoming the first AI DON to support on-chain payments, lazy requests, and unified node discovery. Infernet also becomes the first AI compute mesh to undergo holistic audits; Trail of Bits and Zellic find no critical issues.

We open-source infernet-ml, the first-of-its-kind framework to deploy ML-enabled dApps on-chain. This gives us applied insight into how developers build novel AI experiences, informing our developer experience decisions when building Ritual Chain.

Developers use infernet-ml to bring ONNX & Torch models on-chain, use Stable Diffusion to mint NFTs from prompts, and connect their smart contracts to LLMs.

We come to the conclusion that specialized use cases demand specialized cryptographic gadgets, not general-purpose ones. We begin working on a suite of primitives across provenance, privacy, and computational integrity to give users full flexibility of choice.

These eventually materialize into our suite of modular computational integrity gadgets and verifiable provenance primitives.

Ritual is the only blockchain architecturally flexible enough to let applications choose the primitives that best fit their use case. All other chains must take a one-size-fits-all approach.

We build Frenrug, the first on-chain AI agent, controlled by a combination of LLM and Classical ML models, powered by Infernet, managing >$30,000 on Base.

We are the first to combine text-based transformer models with action-based classifier models to execute on-chain actions. Frenrug informs our design for what Agents should look like.

We publish v0.1.0 of Infernet, the first decentralized oracle network (DON) purpose-built for AI workloads, which can be integrated into any EVM smart contract in under 10 lines of code.

Today, Infernet is the most popular AI DON backed by 9,000+ nodes around the world, used by tens of thousands of users daily, with over 10 million transactions on Base alone.

Ritual emerges from stealth, born from our conviction that AI will completely change how crypto products are used. Our initial architecture outlines a sovereign execution layer purpose-built for AI compute, pioneering a modular precompile and node specialization architecture.

Since then, every chain has converged to ~nearly the same design we pioneered.