# Into: Verathos A Bittensor subnet that makes AI inference prove its own math: every answer ships with a cryptographic receipt that the correct model actually ran. // Inference that proves it ran the real model --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is Verathos? Verathos is a decentralized network for running AI models where each result arrives with a cryptographic proof that the correct model actually did the computation. Instead of trusting a provider's word that you got the model you paid for, you get a mathematical receipt you can check yourself. It runs as Subnet 96 on Bittensor. **The simple version:** It's like getting a tamper-proof receipt with every AI answer, proving the work was done by the model that was promised and not quietly swapped for something smaller and cheaper. **Centralized equivalent:** Think Together AI or Fireworks AI for hosted model inference, but with a verification layer those services do not offer. **How it works:** - **Miners** run AI models on GPUs and generate a cryptographic proof alongside each response. - **Validators** check those proofs on ordinary CPUs in milliseconds, then score each miner on whether the proof holds, plus its speed and throughput. A failed proof zeroes the score for that round. --- ### Why This Matters - **The problem it solves:** When you call a hosted AI API, you usually cannot verify which model actually served your request. A provider can route you to a smaller, cheaper model and you would not know. Verathos makes the model and the computation behind every response checkable. - **The opportunity:** As software agents start spending real money on AI calls, "trust me, it was the right model" stops being good enough. Verifiable inference gives an agent, or the person paying its bills, a way to confirm what it actually bought. - **The Bittensor advantage:** A permissionless set of miners competes to serve requests, and the proof, not a brand name, is what earns trust. Validators do not have to re-run the model on expensive hardware to police the network: they check a cheap proof on a CPU. - **Traction signals:** The verified inference path is live. Verathos exposes an OpenAI-compatible API and ships frequent releases, with recent work adding a Qwen3.6-27B model entry and OpenAI-style tool calling. The verified training half of the system is, by the team's own description, implemented and tested but not yet active on the network. --- ## Full Analysis **Category:** Inference and Compute | **Centralized Competitor:** Together AI, Fireworks AI Hosted AI inference is a crowded market, and most of it runs on trust: you send a prompt to an endpoint and assume the provider served the model it advertised. That assumption breaks quietly when a provider quantizes a model, swaps in a smaller variant, or truncates context to save cost. Verathos exists to remove the assumption, turning "which model ran this?" from a promise into something a client can verify. **Mechanism:** According to the project's repository, any tensor operation, in either inference or training, can be cryptographically proven using sumcheck-based verification over weights that are committed to a Merkle root anchored on-chain. In plain terms: the model's weights are fingerprinted on-chain ahead of time, and each response carries a proof that the computation used those exact weights and produced the output claimed. The proof rests on a few pieces, all described in the repo. A Merkle commitment of the quantized weights lives on-chain, so serving the wrong model is caught. A sumcheck protocol, made non-interactive via the Fiat-Shamir transform, lets a validator confirm the matrix-multiplication results without redoing them. A SHA-256 commitment binds the full output to the proof, so tampering invalidates it. And rather than check every layer, the validator challenges a handful of random layers per request, with detection approaching certainty across repeated queries. On the production side, the proof generator plugs directly into vLLM, a widely used inference server, and produces its proofs in parallel during GPU execution. The repository describes the added cost as single-digit percent overhead, the kind of margin that makes verification practical rather than a research demo. Requests reach miners through an OpenAI-compatible gateway with score-weighted routing, and the same docs note payment options in TAO, USDC on Base, or per-request x402. The roles map cleanly onto Bittensor. Miners need an NVIDIA GPU with 24 GB or more of memory; they serve models and emit proofs. Validators need no GPU at all: they run epoch-based canary tests on roughly 72-minute cycles, verify proofs, and set weights based on throughput, latency, and proof validity. As of the data snapshot, the public repository showed its most recent commit on June 8, 2026, with development carried by a single named contributor across roughly 21 commits. The alpha token traded near 0.01691 TAO with a pool holding about 1,860 TAO in depth, and the subnet was drawing on the order of 2.8% of network emissions at the snapshot. Under Bittensor's current flow-based emission model, that share tracks net staking flows over time rather than any fixed allocation. One note on sourcing: some third-party trackers still label this slot as inactive or reserved. The live on-chain identity and an actively updated repository contradict that, so this writeup follows the repository and on-chain data, which are the primary sources. --- ### Risk Factors - **Single public contributor:** Development is driven by one named contributor in the public repository. That is common for an early subnet, but it concentrates execution and continuity risk in one person. - **Verified training is not live yet:** The inference proof path is in production, but the training proof system, half of the stated vision, is described as implemented and tested yet not active on the network. The full thesis is unproven until it ships. - **Crowded inference market:** Hosted LLM inference is served by many well-funded centralized providers. Verifiability is the differentiator, and it has to matter enough to customers to outweigh the convenience of incumbents. - **Concentration:** A gini coefficient around 0.72 points to moderately concentrated ownership or stake distribution, where large positions can move pool dynamics. --- Into the next one.