GenoVault
v0.1.0 · synthetic data only · measured on a local cluster: one Solana validator, two Arcium ARX nodes
A marketplace where a clinic, biobank or individual puts a dataset to work without handing it over. Runs execute as multi-party computation over ciphertext, the owner's consent is checked on-chain before anything runs, and the encryption key never leaves the owner's browser.
The prototype runs on mock data; the numbers below come from the chain, not from it.
Filled: measured, dated Hollow: not measured yet
Data is secret-shared across MPC nodes; no node sees a record, and the operator holds no key. The key is created and kept in the owner's browser.
Allowed uses, expiry and revocation live on-chain. A run outside them is refused by the program, including one sent straight to the chain, around the API.
The number of records that entered a run is counted inside the computation, not declared by the owner. The split among several owners is not measured yet (M3).
Each figure is read from the chain on a live local cluster, and each check is proven by a negative control on live data: a planted defect the check must catch.
SC-001 · 08 Sep 2026
A canary dataset searched in every account, transaction, log and storage object; 35 328 of 35 328 committed words inspected. The verifier runs without any GenoVault package.
0 on 11
runs · budget 0 on 10
SC-002 · 08 Sep 2026
150 numbers compared, 0 mismatches: the MPC report equals the plaintext recount number for number.
0.0 pp
budget ≤ 2 pp
SC-003 · 08 Sep 2026
Projected from the measured cost of one fold on two corpora; fold cost stays constant within 10 % across run lengths.
This criterion was “≤ 5 minutes”. The measurement found that the cluster saturates at four concurrent computations (×4.08 on sixteen), so 10 000 records take about 55 minutes even with unbounded parallelism. The budget was rewritten after the measurement, not before it.
3.31 h
and 3.71 h
projected · budget ≤ 4.5 h
SC-010 · 08 Sep 2026
The buyer's active time: four reads and one decryption. Wall-clock from order to report was 186 s, on a demo pool of three datasets.
0.1 s
active · budget ≤ 3 min
SC-004 · 27 Sep 2026
10 allowed and 10 forbidden scenarios: 0 false passes and 0 false refusals, on the chain and on the API. Controls: a program without the category check, and the API reading a finalized, 13-second-old chain.
Measured against the buyer category the buyer declares. Verifying that category is planned.
20 of 20
budget 0 false passes, 0 false refusals
SC-005 · 29 Sep 2026
10 revocations. No order was accepted after one; the nearest was refused in the same slot, later in the block. The API sees a revocation within one slot (419 ms). Control: a program without the revocation check accepted 1 587 orders, and the check went red.
0 accepted
after revocation · 10 of 10 · budget < 60 s
Measured on one machine. A production Arcium cluster is a different measurement and has not been made.
SC-006 · M3
Owners' shares, the fee and the refund add up to the deposit, on every run.
SC-007 · M3
A run's fee split among 50 owners in under 30 seconds.
SC-008 · M4
An exported journal reconciled with the chain by a third-party script.
SC-009 · M4
Payment amounts that cannot be read without a viewing key. Today they are public on-chain.
SC-011 · M4
The catalogue's first screen in under 2 seconds.
Prototype
The prototype on this site runs on mock data. There is no API or cluster behind it, and consent is not yet editable from the interface.
No. Every dataset is synthetic, and the system has not been through legal or regulatory review.
The key is created and kept in your browser; the platform never receives it. The nodes compute on secret shares.
The program checks consent before a run exists; without a run, the compute layer has nothing to execute (SC-004, SC-005).
The cluster saturates at four concurrent computations. The time is a measured property of this setup (SC-003).