Developer Documentation

Architectural Overview

Three Main Processes

AIVM executes inference and forwards valid transactions; TheseusStore handles model/context data with DA sampling; HS BFT PoS provides HotStuff-based finality. All communicate via RPC/Networking layer.

Three-Layer Stack

Execution Layer: AIVM

Deterministic tensor-native runtime with Tensor Commits:

Tensor operationsAgent schedulingProof generationSHIP DSL

Availability Layer: TheseusStore

Erasure-coded storage for model weights and agent contexts:

Reed-Solomon encodingRAG contextsStorage minersMerkle roots

Consensus Layer: Proof of Stake

HotStuff BFT with AI-specific requirements:

Valid model rootsVRF selectionOne-block finalityCoupled layers

Block Structure

block.header

header {
  parent_hash, height, timestamp
  post-state Merkle root
  Terkle tree of Tensor Commits
  Merkle root of available model and context blobs
  gas limit
  VRF-selected validator signature
}
body { Transaction[] }

A block cannot finalize unless both conditions hold:

Inference integrity: Every inference must include a valid Tensor Commit proof

Agents availability: Every stored condition must be provably retrievable

TheseusStore Deep Dive

TheseusStore is the availability layer, handling gigabytes of model weights and agent context on-chain.

Model Storage

Immutable weights (Reed-Solomon encoded) addressed by content hash. Cold storage optimized, enforced by miner staking.

Context Storage

Mutable agent data (conversations, embeddings, RAG). Reed-Solomon encoded with faster retrieval.

Agent State Synchronization

Dual-ledger pattern:

On-Chain: Balances, model versions, config in agent state root

Off-Chain: Large context (PDFs, embeddings) in TheseusStore with memory anchors

Sync: libp2p diff-sync with last-write-wins + optional semantic merge

Prover and Verifier Selection

Provers

Run full forward passes. VRF selects by capacity + stake.

Publish hardware specs → Registry tracks → VRF filters (RAM ≥ model) → Cache popular models

Verifiers

All active verifiers check every inference.

Never download weights~2ms per check1K validators → 100 jobs <1s2/3 agreement for finality

Liveness Mathematics

Pr(Liveness) = 1 - (1 - h)ⁿ

With h = 0.33, n = 10 → ≥98% chance at least one honest prover chosen.

Transaction Lifecycle

Model Deployment

Developer uploads weights (with Tensor Commit) to TheseusStore

Inference Transaction

User submits {modelRoot, input, maxGas} to AIVM

Block Proposal

Validator packages model and inference TXs with TheseusStore root

Execution & Proofing

AIVM runs the model and emits a Tensor Commit receipt

Finality

PoS finalizes; TheseusStore miners pin any new context/model shards

AIVM Details →

Deep dive into execution layer

Tensor Commits →

Learn about proof mechanisms

System Architecture