Decentralized machine intelligence has a settlement problem. Networks like Bittensor produce valuable model outputs through a stake-weighted swarm of validators, but those outputs are produced on a foreign chain and are unverifiable to Ethereum, the chain where capital actually moves. Emaxis closes that gap. We specify a four-layer protocol that (1) verifies Bittensor epoch finality on Ethereum with a zero-knowledge light client, (2) proves the correctness of the subnet's model inference with succinct zkML proofs folded into a single recursive commitment, (3) gates access to those signals through a shielded commitment pool that reveals nothing about the subscriber, and (4) lets the verified signal act on the market directly through a Uniswap v4 hook. The result is intelligence that is decentralized in origin, trustless in transit, private in consumption, and economically self-funding at settlement.
The last cycle of crypto x AI produced an enormous amount of unverifiable intelligence. Models ran off-chain, outputs were posted on-chain by a trusted operator, and users were asked to believe two things at once: that the model was good, and that the operator was honest. Emaxis rejects both assumptions. We treat an AI signal as a claim that must be proven before it is priced, and we treat the act of consuming a signal as private information that must be protected before it is acted upon.
Our design target is a single, end-to-end pipeline in which no participant — not the subnet, not a relayer, not the protocol itself — has to be trusted for the system to behave correctly.
Bittensor's value is produced on Subtensor, a Substrate-based chain with its own GRANDPA-style finality and a validator set weighted by TAO stake. Ethereum cannot natively observe Subtensor. The naive bridge is a trusted relayer that simply posts subnet outputs to an Ethereum contract. This reintroduces exactly the trust we set out to remove: the relayer can lie, censor, or front-run.
Three properties must hold for an AI signal to be safely settled on Ethereum:
Emaxis assigns one protocol layer to each property, plus a fourth layer for settlement.
Each arrow is a verification boundary. Data only crosses a boundary if it carries a proof the receiving layer can check on-chain. The layers are independent contracts, composed through narrow interfaces, so any one can be upgraded behind its interface without redeploying the others.
The light client maintains a chain of finalized Subtensor headers on Ethereum. To admit header H_n, a relayer submits a zero-knowledge proof that:
H_n.parent == hash(H_{n-1}) (the canonical continuation of the last finalized header), andH_n was finalized by a stake-weighted supermajority of the committee committed in H_{n-1}.Only after a header is finalized are its committeeRoot (the next signing set) and signalRoot (a Merkle root of the epoch's emitted signals) admitted for downstream use. Individual signals are later opened against signalRoot with a Merkle proof. This removes the trusted relayer: a relayer can submit, but it cannot forge finality.
The coprocessor turns "the subnet says X" into "X is provably the output of model M on input I." It proceeds in four steps per epoch:
output = M(input) for the committed model and a committed input. Verified via the on-chain pairing-based verifier.0x0A) proves a claimed weight or activation against the model commitment.Folding gives Emaxis a property no signal feed has today: the full provenance of every output the protocol ever accepted is verifiable from one 32-byte root, at constant cost.
Verified intelligence is worthless if subscribing to it leaks your edge. Shielded Alpha is a commitment/nullifier pool over a 20-deep incremental Merkle tree. A subscriber:
subscribe(commitment, amount) — deposits the token and inserts a note commitment (shielded membership),claimAccess(...) — spends a note to unlock a signal stream, revealing only a nullifier, the stream id, and a new root,unsubscribe(...) — privately withdraws a note to a fresh address.Identity, amount, and stream selection never appear in calldata. Double-spends are prevented by a nullifier set; a zero-knowledge proof binds each spend to a stream and conserves value. The market sees that someone holds access; it never learns who.
The final layer lets a verified signal act on price directly. Emaxis ships a Uniswap v4 hook with the dynamic-fee permission. On each swap, the hook reads the freshest verified signal from the oracle and prices flow accordingly:
When the subnet predicts benign flow, swaps pay the base fee. When it predicts toxic, adverse flow, the fee ramps toward the cap, protecting liquidity providers from exactly the flow the intelligence flagged. Fees accrue to a treasury that funds the subnet and its validators, so demand for verified intelligence becomes demand for the asset.
The Emaxis token ($EMAXIS) is the unit of access and the flywheel of the system:
| Function | Role |
|---|---|
| Access | Subscriptions to signal streams are paid in $EMAXIS through the shielded pool. |
| Settlement | The v4 hook collects fees that fund the subnet and validators. |
| Alignment | Demand for verified intelligence routes value back to the producers of that intelligence. |
The loop is intentionally circular: better intelligence draws more subscribers and more protected liquidity, which funds more validator stake, which produces better intelligence.
| Phase | Milestone |
|---|---|
| I | Contracts and architecture published; token and v4 hook live on mainnet. |
| II | Production trusted setup; real zkML circuit and KZG model commitments wired. |
| III | Tensor light client finality circuit; trust-minimized relayer set. |
| IV | Shielded subscription mainnet; multi-stream marketplace. |