Skip to content

Repository files navigation

Verathos
Cryptographically verified AI compute on Bittensor

Website · Docs · Try It · Setup Guide


Verathos is a decentralized compute network on Bittensor (Subnet 96) where any tensor operation – in inference or training – can be cryptographically proven via ZK-inspired sumcheck-based verification over Merkle-committed weights, anchored on-chain. Validators verify proofs on CPU in milliseconds and set weights accordingly. The result is a permissionless network where compute is verifiable, not trusted.

Verified Inference – Live

A proof plugin integrates directly into production vLLM serving. It generates sumcheck proofs for GEMM operations in parallel during CUDA graph execution – no challenge-response round trip, single-digit percent overhead. The network exposes an OpenAI-compatible API with score-weighted routing across all miners.

Verified Training – In Development

The same proof system extends to training. The training prover verifies forward pass, backward pass (gradient GEMM), and optimizer step for full fine-tuning and LoRA (AdamW, SGD, Muon). A training job produces proofs that the correct base model was fine-tuned with the claimed data and optimizer. The protocol is implemented and tested but not yet active on the network.

What Gets Proven

Guarantee How
Correct weights Merkle root of quantized weights committed on-chain. Proof checks layer outputs against committed weights. Wrong model = caught.
Correct computation Sumcheck protocol: prover and verifier agree on GEMM results via Fiat-Shamir transform. Covers inference forward pass and training backward pass.
Output integrity SHA-256 commitment over full output, bound to the proof via Fiat-Shamir. Tampering invalidates it.
Probabilistic coverage k random layers challenged per request. Detection approaches 100% over multiple queries.

Architecture

                                  Bittensor EVM
                             ┌─────────────────────┐
                             │  ModelRegistry      │  model specs + Merkle roots
                             │  MinerRegistry      │  endpoints, heartbeats
                             │  PaymentGateway     │  deposits, staking
                             └──────────┬──────────┘
                                        │
        ┌───────────────────────────────┼───────────────────────────────┐
        │                               │                               │
   ┌────▼─────┐                   ┌─────▼──────┐                 ┌──────▼─────┐
   │  Miner   │ ◄── canary ────   │ Validator  │                 │  Gateway   │
   │  (GPU)   │  ── receipt ──►   │  (CPU)     │──shared_state──►│  (API)     │
   │          │                   │            │                 │            │
   │ vLLM +   │ ◄── inference ─────────────────────────────────  │ OpenAI-    │
   │ proofs   │  ── response ─────────────────────────────────►  │ compatible │
   └──────────┘                   └────────────┘                 └────────────┘

Miner – Serves models, generates proofs, registers on Bittensor EVM. Multiple models per hotkey.

Validator – Epoch-based canary testing (~72 min cycles), proof verification, scoring (throughput x latency x proof), weight setting. Proof failure = instant score zero.

Gateway – OpenAI-compatible API. Score-weighted routing to miners. Payments via TAO, USDC on Base, or x402 pay-per-request.

Getting Started

As a User

from openai import OpenAI

client = OpenAI(
    base_url="https://api.verathos.ai/v1",
    api_key="vrt_sk_YOUR_KEY",
)

response = client.chat.completions.create(
    model="auto",
    messages=[{"role": "user", "content": "Hello!"}],
)

Every response includes cryptographic proof metadata. Works with LiteLLM, LangChain, elizaOS, and any OpenAI-compatible client. See the Quickstart and Integrations.

As a Miner

Requires NVIDIA GPU with 24 GB+ VRAM (RTX 4090, A100, H100, etc.).

Quick start (guided wizard handles wallet, registration, funding, HTTPS, PM2):

curl -fsSL https://verathos.ai/install.sh | bash   # or: git clone ... && cd verathos && bash scripts/setup_miner.sh
verathos setup                                       # interactive setup wizard
verathos start                                       # start mining

Manual step-by-step:

git clone https://github.com/verathos-ai/verathos && cd verathos
bash scripts/setup_miner.sh          # creates venv, installs deps, builds CUDA ext

Then create a wallet, register on subnet, fund your EVM address, and start:

python -m neurons.miner \
    --wallet miner --hotkey default \
    --model-id auto \
    --netuid 96 \
    --subtensor-network finney \
    --endpoint https://YOUR-PUBLIC-IP-OR-DOMAIN

--model-id auto detects your GPU and picks the optimal model. See the Setup Guide for wallet creation, EVM funding, model selection, and production deployment with PM2.

As a Validator

No GPU required.

Quick start:

curl -fsSL https://verathos.ai/install.sh | bash -s – --validator   # or: git clone ... && bash scripts/setup_validator.sh
verathos setup validator                                              # interactive setup wizard
verathos start validator                                              # start validating

Manual step-by-step:

git clone https://github.com/verathos-ai/verathos && cd verathos
bash scripts/setup_validator.sh

python -m neurons.validator \
    --wallet validator --hotkey default \
    --netuid 96 \
    --subtensor-network finney

See the Setup Guide for wallet creation, EVM funding, auto-update, and running the gateway proxy.

Repository Structure

verallm/        Verified inference – vLLM proof plugin, chain integration, model registry
neurons/        Bittensor subnet – miner, validator, gateway, scoring, credits
contracts/      Smart contracts (Foundry/Solidity) – UUPS proxies on Bittensor EVM
plugins/        Framework plugins – LiteLLM, LangChain, elizaOS, OpenClaw
scripts/        Setup scripts – setup_miner.sh, setup_validator.sh
examples/       Client examples – OpenAI, streaming, x402
dist/           Pre-built wheels – zkllm and hot-capacity CUDA kernels
docs/           Documentation

zkllm and the hot-capacity audit CUDA runtime are distributed as pre-built wheels in dist/. The setup scripts install them automatically.

Documentation

License

MIT

About

Verified LLM inference and training on Bittensor (Subnet 96)

Resources

Stars

8 stars

Watchers

4 watching

Forks

Releases

Packages

Contributors

Languages