Sirius update We've been heads-down building. Next on the road to mainnet: → Taking training to the next level with confidential computing → Escrow update: pay-per-compute, settled on-chain → Bigger datasets, async training, new models → KYB verification, followed by an independent audit Your data stays protected. The compute comes to it.

Introducing Private AI Training. Your data trains their AI. They never get to see it. Encrypted in your browser → opened only inside a hardware vault → a fingerprint on-chain, never the data → the model ships sealed → one transaction pays the seller and opens the capsule. No way to cheat. Train on data. Without exposing data.

Your data trains their AI. They never get to see it. On Sirius, your dataset is encrypted in your browser before it leaves your laptop. It’s decrypted only inside a hardware vault. Neither we, the buyer, nor the chain can access it. The chain stores a fingerprint. The buyer receives a trained model locked in a sealed capsule. A single transaction pays the seller and unlocks the model simultaneously. No payment without delivery. No delivery without payment. Live on Robinhood Chain testnet. Try it: sirius-data.tech

MPC powered accounts No wallet. No seed phrase. No extension. Your private key is generated as encrypted fragments distributed across a network of nodes and is never assembled in one place. Threshold signatures handle every transaction securely. Sign in, publish a dataset or license one. Same protocol. Same TEE training. Same on chain proofs. The security stays. The friction disappears.

Why does Sirius need a blockchain if the data stays offchain? Because putting private data onchain would defeat the entire point. The chain is the neutral settlement layer: → lock payment → prove settlement → verify identity → record provenance Meanwhile: Data → encrypted & offchain Compute → confidential enclave Settlement → onchain Each layer does one job. No blockchain forced into places where a database works better.

Every dataset on Sirius is anchored on-chain when it’s published. Until today, you had to own it to see that. Now every listed dataset has a public proof page. No account, no wallet : on-chain title, mint transaction, merkle root and the training profile locked at publication. The encrypted file is linked too. Download it, it’s unreadable. A proof only its owner can reach isn’t a proof. It’s a claim.

Sirius now supports binary logistic regression alongside linear regression. Predict continuous values or classify probabilities. The right model for the right problem. More models. More use cases. Same privacy.

Crypto-shredding. Destroy the key → the dataset becomes permanently unrecoverable, even by us. The data disappears. The on-chain proof doesn’t. Verifiable existence. Verifiable destruction.

The borrower-side flow is now live end to end on testnet. Compute travels to the dataset, not the other way around the borrower never touches the raw data. What comes back is the trained model: coefficients, metrics and a verifiable content ID, ready to evaluate against your own test set.

Sirius currently relies on 3 core contracts. SiriusEscrow Atomic settlement between the data owner and user through a SHA-256 hashlock. SiriusKybRegistry Onchain verification of valid KYB status, including expiration and revocation. SiriusDatasetRegistry Records dataset provenance without putting the dataset itself onchain. The blockchain doesn't store private data or run ML workloads. It handles what it's actually good at: settlement, identity & provenance.

The most important part of Sirius doesn't happen onchain. It happens offchain. A dataset goes through: AES-256-GCM encryption client-side Encrypted blob stored on IPFS Decryption only inside a confidential enclave Model training in memory Authorized output leaves, dataset doesn’t The goal is simple: Sirius shouldn't be able to read your data, neither should the buyer. The data doesn't go to the model. The model goes to the data.

Renting data to a stranger creates an obvious problem: Who trusts who first? The buyer doesn't want to pay before receiving the output. The provider doesn't want to provide compute without knowing they'll get paid. Sirius uses a SHA-256 hashlock to bind both events together. The output is encrypted using two components: → a key generated client-side → a secret generated inside the enclave The buyer receives the encrypted output before paying but can't unlock it. The transaction that reveals the secret is also the transaction that settles payment. Payment = delivery. Not two trust-based events. One atomic onchain event.

Today, a data owner basically has 2 options: Keep the data → $0 revenue, asset sits idle Sell the data → lose control the second it's downloaded Neither makes sense. So we're building a third option: don't sell the data, rent the compute around it. More soon

Sirius is entering @RobinhoodApp through @virtuals_io. Sirius is building the confidential data layer for AI. The data stays encrypted. The model comes to the data. The owner stays in control. Private data. Confidential compute. Onchain settlement. CA: 0x72E936815982A577386e8fcDe94f5C9335836C88 app.virtuals.io/virtuals/137…

The world's most valuable data is also the data nobody can use. Healthcare. Finance. Insurance. Industry. Deep history, real-world outcomes, high-quality labels. Exactly what AI needs and exactly what regulation makes nearly impossible to share. Our thesis with Sirius is simple: What if you could extract value from a dataset without ever sharing it? That's what we're building sirius-whitepaper.vercel.app