Sirius / Whitepaper Where we are

Whitepaper · August 2026

Sirius

Rent out valuable data without ever exposing it.

The dataset stays encrypted and under its owner's control. The borrower submits a training job that runs inside a hardware enclave and gets back only the trained model. One transaction pays the provider and delivers the result at the same instant.

3 contracts, deployed and verified 0 chain-specific code inside the enclave Intel TDX enclave Non-custodial
Scroll

Rent out valuable data without ever exposing it

A data owner has two bad options today: keep the data and earn nothing, or sell it and lose control forever. Sirius opens a third one — rent it. The encrypted dataset never leaves the owner's control. The borrower submits a training job that runs inside a hardware enclave and gets back only the trained model. Settlement and audit happen on Robinhood Chain.

3
contracts deployed and source-verified on Robinhood Chain, auditable without asking us
0
chain-specific code inside the enclave — 721 lines, none of them about settlement
0.12 $/h
confidential compute cost (Phala tdx.medium CPU) — the enclave is not a structural cost

What makes it work

The computation moves, not the data
The dataset is encrypted in the owner's browser, stored on IPFS, and only ever decrypted inside an Intel TDX enclave. Nobody — not the borrower, not Sirius, not the cloud operator — reads the raw rows.
Payment and delivery are the same on-chain event
The model key is locked under two factors at once: a non-exportable browser key and the preimage that settlement publishes on-chain. One transaction pays the provider and publishes the secret that unlocks the model. Neither party can cheat the other.
Sirius holds nothing
No funds, no dataset token, no plaintext. Funds live in the user's wallet or in the on-chain escrow. In enclave mode, Sirius cannot decrypt a dataset even if fully compromised.

The most useful data is the data nobody shares

Health records, transaction histories, industrial telemetry — the datasets with the highest value are the ones under the tightest legal and commercial constraints. They are never shared in the clear, so they stay idle.

Why the existing answers fail

ApproachWhy it breaks
Sell the datasetControl is lost at the first download. Data copies perfectly and forever.
Contracts and NDAsLegal recourse after the fact, not technical prevention. You find out once the damage is done.
WatermarkingA perturbation small enough to preserve utility is small enough to be stripped or lost in noise.
A trusted intermediaryRecreates exactly the risk you were trying to remove, and adds a rent.
Sirius inverts the problem: the data does not travel to the model, the model travels to the data — inside an enclave where nobody, Sirius included, can read it.

Five steps, from upload to settlement

A full cycle: the provider publishes a dataset, a borrower locks funds and submits a job, the enclave trains and produces proof, and one transaction closes the trade.

The raw data never leaves the provider → IPFS → enclave axis. Only the model and the proofs travel.
01
Deposit
The provider's browser encrypts the CSV (AES-256-GCM) for the runner's ingestion key. The enclave opens it, computes a Merkle root and public metrics, re-encrypts it under a random per-dataset key, and pins the blob to IPFS. The server never sees plaintext.
02
Tokenize
A dataset title is minted on-chain carrying the dataset's identity — IPFS fingerprint, Merkle root, metrics. A KYB credential attests the entity. The token is the title; the data stays encrypted off-chain.
03
Borrow
The borrower browses public metrics and a sample, then locks funds against a hash whose preimage is held only by the enclave, plus a deadline after which the funds return.
04
Train
The runner fetches the encrypted blob, unwraps the key with the enclave-sealed master key, verifies integrity, decrypts in memory, trains, passes the result through an output gate, and produces a TDX attestation bound to the result.
05
Settle
The borrower stores a locked capsule containing the model key. The runner submits EscrowFinish: the provider is paid and the fulfillment becomes public in the same transaction, which is what opens the capsule.

Who can read what

ArtifactProviderBorrowerSiriusEnclave
Raw datayes (their own)nonoyes
Encrypted blob on IPFSyesyesyesyes
Decryption keynononoyes
Public metrics and sampleyesyesyesyes
Trained modelnoyes, after paymentnoyes
Escrow fulfillmentnoonly once published on-chainnoyes

Payment and delivery become one event

This is the core of the protocol. A time-based escrow only protects the buyer. Here the release is controlled by neither party — the secret that unlocks the model is held by the enclave and revealed only by the transaction that pays.

The double lock

After a successful job, the runner encrypts the model key under two factors that must both be present. The first is a private key generated in the borrower's browser, non-exportable and never transmitted — so nobody else can ever open the capsule. The second is the escrow fulfillment, which exists only inside the enclave. The borrower receives and stores this capsule before any payment. It is useless until the fulfillment is published, and the only thing that publishes it is the EscrowFinish transaction that pays the provider.

Two possible outcomes, neither of which leaves a party exposed.

Every outcome

ScenarioResult
Everything worksOne transaction pays the provider and unlocks the capsule the borrower already holds.
Dataset is unusableThe job fails, no fulfillment is ever produced, CancelAfter refunds the borrower.
Borrower tries to get the model without payingImpossible. The capsule needs a secret that only the payment transaction publishes.
Sirius is compromisedThe fulfillment and the master key live in the enclave. A compromised server can stall the trade, not steal from it.
Dispute over data qualityNot resolved by escrow alone. Challenge period and arbitration are on the roadmap.
Once created, the escrow can only resolve two ways. The fulfillment is never stored — it is regenerated at release time.

Settlement where the assets are

Robinhood Chain is an Arbitrum Orbit L2 that settles on Ethereum, live on mainnet since 1 July 2026, with a public permissionless testnet open since February. Gas is paid in ETH, contract deployment needs no allowlist and no KYC, and the chain exists to carry tokenised real-world assets — which is the market Sirius sells into.

What the chain gives us

A secret can be published atomically with a payment
This is the single property fair exchange requires, and the only thing Sirius asks of a settlement layer. A hashlock contract provides it: the transaction that pays the provider is the transaction that reveals the preimage.
The secret becomes state, not history
Storing the preimage costs 20,000 gas, and lets a browser that kept nothing but its loan id reopen its capsule months later. A ledger that deletes its escrow object on completion cannot offer this.
Fees can be sponsored
All three ERC-4337 EntryPoints are deployed, so a data provider can list an asset without ever holding crypto. For a hospital or a bank, that removes the first and largest obstacle.
Front-running is neutralised by the chain
Ordering is strictly first-come, first-served with no priority fee, so the canonical attack against a hashlock contract cannot be bought.

What we accept in exchange

Stating this plainly is more useful than a case that only favours us.

Finality is soft for roughly thirteen minutes
The chain confirms in under a second, but stays revocable until its batch reaches Ethereum. There is a window in which a model can be delivered before payment is irreversible. This is the most serious trade-off and it is irreducible.
The escrow is code we own
A hostile payee that reverts on receipt could otherwise block the transaction, so settlement is pull-only and the release path makes no external call. That is a correctness argument we defend ourselves, and it is why the contracts are small and verified.
On-chain records are permanent
A Solidity event is eternal and ends up posted to Ethereum. The dataset registry therefore records only opaque digests — no name, no description — and the right to erasure remains served by destroying the decryption key, never by an on-chain operation.
Decentralisation is early
A single sequencer, two whitelisted validators, and a Stage 0 rating. Force-inclusion via L1 is not a reliable escape hatch on this chain today.
On portability The enclave holds no chain-specific code. Its 721 lines cover sealed key derivation, attestation and the output gate, and none of them mention a settlement layer. Everything chain-specific lives in one replaceable adapter, which is a deliberate structural choice rather than a claim about any particular chain.

Three contracts, deployed and verified

Three contracts carry the protocol on-chain. Each is deliberately small, has no administrator on the settlement path, and has its source verified on the explorer — so it can be audited without asking us.

ContractRoleWhat it does
SiriusEscrowSettlementSHA-256 hashlock escrow. Publishing the preimage and entitling the provider happen in one transaction that makes no external call at all.
SiriusKybRegistryIdentityPer-entity KYB with explicit subject consent, revocation, expiry and multiple verifiers.
SiriusDatasetRegistryProvenanceNon-transferable dataset title with a deterministic identifier, and a tombstone rather than deletion after crypto-shredding.

Live on testnet, chain 46630

SiriusEscrow · 0xf2f8f5ba5c16c780179af0da2f250deceaf0064f
SiriusKybRegistry · 0xc88346fc152c1a2de065bd9a29c18321db8abb01
SiriusDatasetRegistry · 0x7d9efc6abdbb6d898aea4ce775ca95b483dffdb8

Source code is verified on Blockscout for all three. Total deployment cost was 3,497,006 gas. A full escrow cycle — lock, release, payout — costs roughly 0.000005 ETH and has been executed against these contracts on the live chain.

The gap in on-chain compliance

Robinhood Chain applies compliance controls at the network layer, and they work. They are also unavailable to the applications running on it — which is where a registry belongs, and where we happen to have built one. This section states exactly what we checked, and where our claim stops.

What exists, and what does not

Transactions from sanctioned addresses are filtered at the sequencer, and an analytics partner provides monitoring on top. Those controls are part of why the chain is credible for regulated assets.

They are also opaque to applications. A contract cannot query them. A dApp has no way to ask, on-chain, whether the counterparty it is about to serve is a verified business, and therefore no way to build its own logic on top.

We checked for the standard building block: EAS, the ecosystem's attestation layer, is not deployed at its canonical addresses, and the chain's documented protocol contracts include no identity or registry contract of any kind.

This produces a real tension on a chain built for regulated finance: anyone can deploy a contract without any KYC whatsoever, while the assets being tokenised are securities.

Why ours generalises

SiriusKybRegistry was written for our own gating requirement. It attests that an address belongs to a verified business, requires the subject's explicit consent, supports revocation and expiry, and allows several independent verifiers under a two-step admin.

Nothing about it is specific to data lending. Any application on the chain could read isKybValid(address) and gate on the result.

Scope and method We are not claiming to have solved compliance for this chain, and the claim above is deliberately narrow. Here is exactly what it rests on. We queried the three canonical EAS addresses on chain 46630 and found no deployed code at any of them. We read the chain's documented protocol contracts and found no identity registry, no allowlist and no attestation contract, and no third-party registry that another contract could call.

Network-level filtering and off-chain monitoring do exist, and we are not disputing them. What is missing is more specific: none of it is callable from a contract, which is what an application needs in order to gate its own logic.

Nor is this an argument for identity at the protocol level. Permissionless deployment is what draws builders to a chain, and a registry that everyone must pass through would remove it. Ours is opt-in: a contract other applications may call when their own counterparties require it, and ignore entirely when they do not. If a composable registry already exists or is planned, we would rather build on it than beside it.

Four layers, one trust boundary

The browser, the application server and the database are all treated as untrusted. Only the runner inside the confidential VM sits within the trust boundary, and it is deliberately small.

The browser, the server and the database are all treated as untrusted.
One sealed root key
In enclave mode the master key is derived inside the confidential VM by the dstack KMS, released only to a measured application identity. Every downstream secret descends from it: the per-dataset key, the model key, the receipt HMAC, and the escrow fulfillment. Sealing one source closes the whole derivation graph at once.
A small, stable measurement
Rather than shipping the whole web application into the enclave, only the compute core runs there. The measured code stays small and stable, so a frontend change does not invalidate the attestation, and the database stays outside where it belongs.
Authorization delegated by the wallet
One wallet signature delegates a non-exportable browser key for an hour. Every sensitive operation then carries a 60-second grant scoped to the operation, the wallet, the ids and a nonce. The transport secret between server and runner grants no business rights at all.
Verifiable attestation
The runner produces an Intel TDX quote bound to the result hash. Before accepting a result and authorizing settlement, the server verifies the binding, the Intel signature chain and a pinned measurement. The runner never verifies its own quote — that would be circular.
Sealing one root closes the whole derivation graph at once, rather than securing each use one by one.
An honest distinction A valid quote proves that genuine Intel hardware signed the result. It does not by itself prove that the Sirius code produced it — that requires verifying the code measurement, which is still open work. Until it lands, that specific guarantee rests on trusting Sirius as an operator.

A working application, screen by screen

Not a mockup. The full path — deposit data, tokenize it, borrow it, train in the enclave, retrieve the model — is navigable end to end today.

sirius.xyz
Sirius
Sign in

Sirius

Your data, put to work without ever exposing it.

Connect a wallet
Read the docs
Landing — a particle blob unfolds as you scroll.
Dashboard
Balance read live from the ledger, card on-ramp for users who have never touched crypto, and the three things you can do: browse the catalog, train, or publish data.
Marketplace
Public volumes per dataset — row and column counts only, with no per-column statistics, no schema and no raw rows. Borrowing signs the escrow client-side — the funds never pass through Sirius.
Train
Self-training on your own data and borrowed jobs share one queue. A settled loan expands to show the model, its metric, and the settlement transaction.
My data assets
Visibility is one click and reversible. Deletion destroys the decryption key — crypto-shredding, irreversible by design, which is the technical answer to an erasure request.
Wallet
Balance, on-ramp, and an explicit way out: send to any address you control. Funds are never trapped inside the product.
sirius.xyz/dashboard
Sirius
DashboardTrainMarketplaceMy dataWallet
Dashboard
0x20f8e8…023CA0
Balance
0.42ETH
Add funds
MarketplaceBrowse datasets and start a confidential training run.
TrainTrain on your own data or on a dataset from the catalog.
My dataDeposit, tokenize and monetize a dataset.
Dashboard — balance read live from the ledger, card on-ramp, and the three things you can do.
sirius.xyz/marketplace
Sirius
DashboardTrainMarketplaceMy dataWallet
Available datasets
Borrow access through a conditional escrow. The data stays encrypted — you only get back a model trained inside the enclave.
Care pathways — 2024 cohort
★Listed
48,210 rows17 columns6.4 MB
agestay_daysreadmitscore
67400.82
54910.41
73200.95
0.004ETH
Borrow
Marketplace — public metrics and a sample. Borrowing signs the escrow client-side; funds never pass through Sirius.
sirius.xyz/train
Sirius
DashboardTrainMarketplaceMy dataWallet
Train a model
The computation runs inside the enclave — you get the model, never the raw data.
My training runs
Care pathways — 2024 cohort
Borrowed
SETTLED
E-commerce transactions 2025
Borrowed
TRAINING
Telecom churn (internal)
Self-train
DONE
Model delivered
Model (CID)
bafkrei…u7qd
R²
0.8413
Settlement tx
A19F…2C7B
Decryption key
••••••••••••
Train — self-training and borrowed jobs share one queue. A settled loan expands to the model, its metric and the settlement transaction.
sirius.xyz/datasets
Sirius
DashboardTrainMarketplaceMy dataWallet
My data assets
3 datasets · 2 public. Each token is the on-chain title to your data.
Care pathways — 2024 cohort
Public
48,210 rows17 columns6.4 MB
CID   bafybeih…7k2q
Merkle   4f1c…9ae0
Token   00000BC8…
Public Unlisted Private
Delete
My data assets — visibility is one click and reversible. Deleting destroys the decryption key: irreversible by design.

The design stance

Technical depth is optional. A Simple/Advanced toggle hides transaction hashes, CIDs and Merkle roots by default, and appears only where it changes something. First-time visitors are detected by reading the ledger rather than writing to a database. Sign-in offers a Google login backed by MPC key shares before it offers an external wallet — the target user is a data owner in a hospital or a bank, not a crypto native.

What runs, and what does not yet

The settlement layer is deployed and exercised on-chain. The application is not yet wired to it, and the enclave has never run on real hardware. Those two gaps matter more than anything else in this document, so they are stated before the rest.

On-chain

ComponentStatusEvidence
Escrow, KYB registry and dataset registryShippedDeployed and source-verified on testnet 46630 · 49 contract tests
Fair-exchange cycle on the live chainShippedLock, release and payout executed against the deployed contracts
Confidential core adapted — chain-bound key derivation, versioned capsuleCode complete84 unit tests; not yet exercised end to end
Runner operations against the contractsCode completeWired to the deployed addresses
Training inside the on-chain cycleNot demonstratedThe validated cycle covers settlement only
Application wired to the contractsNot startedInterface, database and orchestrators still target the previous rail

Product

ComponentStatusEvidence
Provider pipeline — encrypt, IPFS, Merkle, dataset title, KYBShippedValidated end to end on testnet
Marketplace and conditional escrowShippedValidated end to end on testnet
Confidential training and atomic settlementShippedRun, release and model decryption validated on testnet
Non-custodial embedded wallet, API authentication, fiat on-rampShippedClient-signed escrow, MPC key shares, no key export
Per-dataset keys, crypto-shredding, visibility controlsShippedCore validated; on-chain destroy still to confirm

Enclave

ComponentStatusEvidence
Enclave-sealed master key and TDX attestationCode completeValidated against a real quote fixture and a local simulator only
Isolated confidential runner as an HTTP serviceCode completeWired end to end; not yet deployed to a real CVM
Deployment to a real Phala CVMNot startedThe decisive milestone
Full code-identity verificationNot startedRequires a real CVM
Read this before anything else No TDX quote has ever been produced by real Intel hardware. Verification was validated against an authentic quote fixture and a local simulator. That is a difference in kind, not in degree, and it is precisely what the CVM deployment must resolve.

Current capability is deliberately narrow: deterministic linear regression on CSV, files up to 50 MB, synchronous jobs with a 60-second timeout, CPU only. Arbitrary borrower code and long training runs are post-MVP and will require an async job queue, sandboxing and differential privacy before they can be safe.

What comes next, and why it matters

Five phases. The first is not optional — until it lands, the central claim of the protocol is an assertion rather than a proof.

01
Real enclave
Docker packaging, deployment to a stable confidential VM, the first quote produced by real hardware, pinned measurement, full code-identity verification. Critical checkpoint: key persistence across redeployments must be proven before any real data is encrypted, or a redeploy makes datasets unrecoverable.
02
Application on the EVM rail
Wiring the interface, database and orchestrators to the deployed contracts, so the product runs on the same chain it settles on. Sign-in moves to wallet signatures supported by every vendor, which removes a defect where only one wallet could complete authentication. Estimated at three weeks; it is plumbing, not research.
03
Trust and compliance
On-chain reputation computed from ledger events, dispute arbitration through the cancellation window, and a public audit page. External legal opinion on erasure, on data-intermediary status, and on the status of a model that may memorize.
04
Wider workloads
Async job queue, longer training runs, differential privacy with a real privacy budget, and provider-set sensitivity levels. This is what unlocks borrower-supplied code, which the current output gate cannot yet make safe.
05
The AI layer
A discovery agent over public metadata, then a synthetic data generator inside the enclave — the one genuine research effort, because its quality decides how freely a data scientist can work without ever seeing a real row. Then multi-party dataset joins.

Two people, and a verifiable execution trail

Ali Ben Yezza on business and partnerships, Noé Wales on the protocol, out of the DeVinci Blockchain association. Small team, narrow scope, documented decisions.

The repository carries 27 dated architecture decisions, each with its rationale and the alternatives that were rejected. Every increment ends with a security review whose findings are written down and closed in the next one. Where the code is more honest than a pitch would be, the code wins: the rate limiter is commented in the source as friction rather than a hard bound, and this document says the same thing.

What we are looking for

Pilot data providers
One organization in health, insurance or finance holding a dataset it cannot sell but would rent under an enclave. This is the scarcest input and it unlocks everything else.
Two hires
A cryptography and systems-security profile to remove the single-person dependency on the confidential core, and an enterprise sales profile to carry sales cycles that two founders cannot run alongside the product.
A compliance partner
Counsel able to qualify crypto-shredding as erasure and to settle our status as a data intermediary. We would rather have an unfavorable answer now than a surprise in two years.
On tokens There is none and there will not be one. Revenue is in fiat or stablecoin, no issuance is planned, and the protocol depends on no asset whose supply we control. Robinhood Chain is a settlement and audit rail, not a fundraising vehicle.