Verifiable data, from databases to agents
Provably correct, not just probably right.
Blockchains, databases and AI agents pass data to each other that nobody can check. We attach a proof to every answer, so anyone can verify it in milliseconds.
Verifiable Databases
SELECT MAX(price) FROM eth_usdc_trades WHERE block_number >= 21433523;
ETH / USDC, 7-day high
$3,219.77
proof 1.1 KB
from 1,000,000 rows
from 1,000,000 rows
Investors and partners
Proofs wherever data moves.
Each one proves a different kind of answer: a query result, an agent's tool call, a reply from another company's agent.
The research behind it.
Accepted into ACM CCS 2026
QEDB: Expressive and Modular Verifiable Databases
Every query comes back with a proof that the answer is correct, so your database can run on infrastructure you don't control.
Read the paper~1 KB
Proof size on most queries, whatever the number of rows
<25 ms
To verify aggregation and selection queries
No SNARKs
Accumulators and vector commitments. No circuits to write.
All papers, newest first
The people building it.
Backgrounds in cryptography, AI and distributed systems, from IBM, Intel and university research.
Meet the team
Emanuele Ragnoli
CTO and founder
Inventor and adjunct professor. Previously at Intel and IBM Research.
Alberto Trombetta
Scientific officer and founder
Professor of computer science, University of Insubria.
Simone Bottoni
Engineer and founder
PhD in secure, privacy-preserving techniques for federated machine learning.
Advisors
Benedek OrbanCV Labs
Matteo CampanelliOffchain Labs
Nicolas MohnblattZK Security
From the blog.
All posts- SourceryKit · 8 min readThe Evidence Was Recovered. Did It Reach the Answer?In 9 of 24 repaired runs, no model was asked to write a new answer after recovery. Where synthesis did run, an execution ledger raised supported values reaching the answer from 73.9% to 81.2%, a direction but not yet a significant effect.
- SourceryKit · 10 min readThe Trace Was There. The Retry Was Blind.On the same 60 failed runs, SourceryKit-guided recovery restored complete tool evidence in 53.3%, compared with 30.0% for blind retry. But repairing the evidence did not always repair the final answer.
- SourceryKit · 9 min readThe Agent Was Right. The Evidence Wasn’t.We benchmarked 15 leading AI models and found SourceryKit detected 100% of the covered deterministic errors—including 109 unsupported claims that answer-only scoring accepted.
