The Verified Inference Network
Mantissa pools consumer GPUs into a permissionless network where every LLM response ships with a falsifiable receipt. Its Exact Inference Profile makes execution bit-identical across the validated NVIDIA and AMD paths. Verification becomes a hash comparison, not a judgment call.
SHA-256(logits) = 04436b5a…, identical on NVIDIA·CUDA, NVIDIA·Vulkan, and AMD·Vulkan.
The Problem
Consumer GPUs spend substantial time idle. What has kept them from becoming open AI infrastructure is not the hardware. It is the trust gap: you cannot see whether an anonymous machine really ran the model you paid for.
An anonymous operator's most profitable strategy is to serve something cheaper than promised: a 4-bit quantization sold as 8-bit or a 3B model sold as an 8B. The output can still look plausible. Any open network that pays for inference must first answer one question: did this machine really run this model on this input?
Floating-point addition is not associative, so identical computations produce different low-order bits on different silicon. Because honest variation looks like misbehavior, every prior verification approach paid for the ambiguity through replication, tolerance bands, or proof systems costing thousands of times the computation itself. Mantissa removed the ambiguity instead.
The Breakthrough
The Exact Inference Profile (XIP) is a constrained execution mode under which every conforming device produces byte-identical outputs at every token position, regardless of vendor. We validated it end to end on real NVIDIA and AMD hardware.
Different hardware. Same bytes. That turns a disputed inference from an argument into a computation anyone can repeat.
| Model class | Validated on | Result under XIP |
|---|---|---|
| Autoregressive dense | Qwen2.5 0.5B and 3B (Q8_0) | Byte-identical logits at every position, across all three backends, prefill and decode |
| Diffusion dense | LLaDA-8B (Q8_0) | Cross-vendor score divergence exactly 0.000 on every sweep prompt |
| Diffusion mixture-of-experts | LLaDA-MoE-7B (64 experts, Q8_0) | Deterministic expert routing; cross-vendor divergence exactly 0.000 on every prompt |
How It Works
With bit-exactness, verification stops being statistics and becomes arithmetic. Payment stops being trust and becomes settlement of falsifiable claims.
Every job produces a signed receipt: model hash, input commitment, output commitment, and intermediate checkpoint hashes. It is a falsifiable claim: this model, this input, these exact bytes out.
One executor produces the result; a sampled fraction of receipts is checked by independently drawn conforming auditors. Exact matches pass, while any mismatch escalates without making every job run k times.
Receipts batch into Merkle rollups committed per epoch to Solana, with an optimistic dispute window before finality. No custom blockchain.
A challenged receipt escalates to independently assigned conforming referees, which re-execute the disputed work and compare exact hashes. Truth is a reproducible computation, not a vote.
The Network
Mantissa coordinates geographically distributed GPUs without relaxing what correct execution means. Hardware varies. The receipt does not.
The Mantissa Stack
Execution, verification, and settlement are separated behind stable interfaces so each layer can evolve without compromising receipt integrity.
A memory-safe Rust client built on libp2p for identity, discovery, and gossip. Nodes join only after a conformance self-test proves their hardware reproduces the profile bit-for-bit.
XIP exact-hash verification with sampled teacher-forced audits as the primary mechanism. Verifier assignment is drawn by verifiable randomness from the stake-anchored registry.
Signed receipts batch per epoch into Merkle rollups committed to Solana with an optimistic dispute window. Mantissa runs no blockchain of its own.
Measured, Not Promised
Every figure below comes from the project's physical two-vendor validation rig: one NVIDIA GPU, one AMD GPU, three backends, raw 32-bit logits hashed at every position. Evidence records are retained for the technical publication.
What It Is For
We publish the unfavorable number ourselves: a verified consumer-GPU node cannot beat centralized commodity inference on price. The product is the receipt, built for workloads where an AI's output carries authority and must be auditable.
Autonomous agents move money, sign transactions, and modify production systems. Mantissa gives every step a receipt: exactly this model, exactly this input, exactly this output. Anyone can check it and challenge it on-chain.
An AI-assisted decision that may face challenge in compliance, adjudication support, or published research can be re-verified bit-for-bit years later, by a regulator or a court, on any conforming hardware.
Cross-vendor bit-exactness gives scientific and safety-critical users identical results on conforming GPUs. Cross-vendor redundancy can expose silicon and driver faults that homogeneous fleets may share.
Roadmap
Every widening of scope ships with its own cross-vendor evidence, and the launch review is an explicit human decision gate. We publish the gates, not promises about clearing them.
Cross-vendor bit-exact inference validated for dense, diffusion, and diffusion-MoE models; receipts, sampled audits, and referee settlement proven live on two-vendor hardware and the simnet.
Finish the signed Windows operator app and enrollment path, then run clean-install friend testing across real operators. The headless Rust loop and signed engine payload already exist. In parallel, XIP productization is recovering deterministic prefill parallelism and safe fast paths. Every optimization must pass the full cross-vendor conformance matrix.
Open node onboarding behind the conformance gate, epoch receipt rollups and the dispute path on Solana testnet, and public network telemetry.
Developer gateway and agent features, the launch review as an explicit decision gate, then the capability ladder: metro clusters for 70B-class models, with wide-area frontier sharding as a gated research track.
The Thesis
The whitepaper covers the root cause, the exactness result, the receipt protocol, the token economy, and, deliberately, the limits. Credibility requires stating both the result and its boundaries.
Decentralized GPU markets have always failed at the hardest question: proving an anonymous machine really ran the promised model. Mantissa answers by removing the root cause. Under the Exact Inference Profile, conforming execution is bit-identical across the validated hardware, so an independent audit can establish an exact match or mismatch.
The network launches with whole-model nodes for practical open-weight models, then scales through metro clusters and versioned conformance profiles.
Developers buy inference credits by burning $XIP. Operators mint only after epoch-settled receipts survive the dispute window; supply is tied to verified demand, never to hardware presence. Mining without verified demand mints nothing.
Stake gates throughput, reputation buys priority within safe bounds, never security. Confirmed misbehavior is slashed.
Each receipt commits to model hash, input commitment, output commitment, and checkpoint hashes. Auditors are assigned by verifiable randomness and re-execute sampled claims on conforming hardware.
Under exactness, disputes collapse to byte comparison: a challenged receipt escalates to independently assigned conforming referees for exact re-execution.
mantissa.network
Route agent actions, research pipelines, and high-assurance inference through an API designed to return evidence with output, or put your GPU to verified work.