Nibiru Engine
Private execution. Checkable receipts.
Nibiru is our verifiable confidential compute engine. It combines execution on encrypted data with receipts for a committed computation, so a verifier can check a declared claim without opening the private witness.
For work that cannot simply be handed over.
A data owner may need an outside operator to evaluate a private policy or model. The owner needs confidentiality; the recipient needs evidence of which computation ran. Nibiru develops those two capabilities together, starting with bounded workloads.
How the pieces fit
- 1
Define the job. Commit to the program and the relevant input and state identities under a supported execution profile.
- 2
Execute on ciphertext. Fully homomorphic encryption (FHE) supplies the private execution layer.
- 3
Produce the semantic receipt. Zero-knowledge (ZK) proofs check the declared computation over commitments. Multiparty computation (MPC) supports mutually private witness generation in the applicable modes.
- 4
Verify the statement. Check the receipt against the intended program, artifact, and verifier profile before using its result.
What a receipt establishes
The receipt statement defines the program and state commitments, semantic relation, and artifact bindings it checks. Different modes support different statements and privacy boundaries.
The semantic proof does not itself prove the encryption implementation or authenticate arbitrary source data. The execution-to-receipt binding, cryptographic assumptions, and verifier configuration are part of the profile a deployment must qualify.
Current scope
Implemented + Verified finite: recorded encrypted-execution, semantic-receipt, and recursive smoke profiles. Core bounded-computation semantics and compilation results are machine-checked in Lean 4 at their exact theorem scopes.
Open: broader workload and performance qualification, remaining cryptographic and runtime bindings, and independent production qualification. Timings from one retained profile do not predict another workload.
The public viewer includes a guided simulation and a separate retained run record. It lets you inspect the public proof surface; it does not submit a live encrypted job. See the status definitions for how we describe these results.