Verathos (SN 96) Lays Out a Five-Phase Roadmap to Replace Trust With Cryptographic Proof

Verathos, the verified AI compute network on Bittensor Subnet 96, recently laid out a five-phase roadmap framing its entire build around replacing trust in a provider's word with mathematical proof. The team says each phase builds on the one before it, moving control away from any single provider and toward proof-based computation.

"Verathos exists because trust isn't good enough, when a request goes out to a model, you get results back but you simply have to trust that the right model ran and the provider is honest."

Replacing Trust With Math

Phase one, verified inference, is live today.

Every inference on the network returns a cryptographic proof, miners and models register on chain, and onboarding runs without a gatekeeper. Verathos also put a mesh operator system into testnet this month, letting miners pool independent hardware, eight machines combining 820GB of VRAM across two active meshes running GLM-5.2 and Qwen3.6-35B-A3B, to serve large open models under this same verified framework.

Phase two, intelligent routing, is in active development. Verathos currently routes requests to miners on a score-weighted basis, and the team says the next version sends each query to the model built for it, turning the network into a set of specialists instead of one generalist trying to cover everything.

Phase three, verified action inference, is already in testing. The proof system built for LLM inference extends to action models powering robotics, checking whether a model produced the correct action under the same cryptographic guarantee used for text. The team says this phase originally scoped fine-tuning for domain-specialized experts before widening into verifying model-produced actions more broadly.

Phase four, agent-orchestrated experts, is planned next. Once routing and verified action inference both ship, the team says the network coordinates multiple models on a single task instead of handing it to one, and adds research agents flagging gaps in the network's knowledge and surfacing what to train next.

Phase five, autonomous intelligence, is the long-term target. Verathos describes continuous learning, multimodal experts, and persistent memory models that improve their own capabilities over time instead of only answering the question in front of them.

How Verification Works Under the Hood

The proof system underpinning all five phases works the same way today as it will at phase five. The network uses sumcheck-based cryptographic proofs over Merkle-committed model weights, and validators check those proofs on CPU in milliseconds, with overhead in the single-digit percentage range.

A proof plugin integrates directly into production vLLM serving, generating proofs during CUDA graph execution, and a gateway routes requests to miners by score while accepting payment in TAO, USDC on Base, or per-request through the x402 protocol.

Phase one is live today. Test it yourself by pointing a request at the Verathos OpenAI-compatible API, which returns a cryptographic proof alongside the answer. Watch phase two next; intelligent routing decides whether Verathos becomes a smarter network or only a faster one.