Skip to main content

​
Confidential VM Architecture

TL;DRDarknyx matches hidden orders inside an Intel TDX confidential VM. Hardware attestation lets a client identify the measured software before revealing an order, while Solana accepts settlement only from the registered signer set and only with a valid zero-knowledge proof. This reduces trust in the infrastructure operator, but does not erase it: clients still trust the approved measurement, Intel TDX, governance, and the matching policy implemented by that measured code.

​
Why matching needs a confidential computer

A central limit order book normally needs to see every order. Publishing those orders on-chain would reveal price limits, sizes, timing, and strategy before execution. Pure zero-knowledge matching can hide more of that trust boundary, but makes a feature-rich, low-latency book substantially harder to operate. Darknyx uses a confidential VM as the private execution environment:
  • encrypted memory limits what the host and infrastructure operator can inspect;
  • a hardware-signed quote identifies the software measurement and boot state;
  • darknyx-tee terminates TLS itself with a boot-random key. Clients bind that certificate to a fresh TDX quote, while the dstack s-suffix route passes the encrypted stream through without learning the order intent (see Transport & Attestation);
  • the service derives an ordered set of settlement signers, one per Merkle shard.
This preserves the familiar order-entry experience while keeping custody and settlement validity on Solana.

​
From measurement to an accepted settlement

Attestation is useful only when it is connected to authority. Darknyx closes that loop in three checks:
The quote binds a fresh client challenge and a hash of the complete ordered signer set. The client also compares those signers with finalized on-chain configuration. Checking only the hostname, the self-reported compose hash, or the shard-0 key is insufficient.

​
Which layer guarantees what

This separation is intentional. Asset validity is proof-enforced; matching policy remains an attested-code guarantee.

​
What an infrastructure operator can still do

With an approved, correctly attested image and an uncompromised TDX platform, the host should not be able to read plaintext order memory or forge a proof-valid asset transition. It can still:
  • stop or restart the service;
  • delay, drop, or censor network traffic;
  • observe timing, volume, and other network metadata;
  • attempt to deploy a different image, which clients must reject unless it is an approved release and governance has registered its signer set.
A flaw in the measured application, the hardware, the attestation verifier, or the governance process can weaken these properties. Darknyx therefore treats quote verification, finalized key refresh, independent circuit review, and split multisig control as core parts of the operating trust model.

​
Why one CVM, not a committee

A committee can distribute trust, but adds coordination, latency, and a collusion threshold. Darknyx currently chooses one confidential matching service with multiple shard signers derived inside it. That service can host several isolated market books behind the same attested endpoint. The product tradeoff is straightforward:
  • faster private matching and a simpler client protocol;
  • one attested software release and signer set to evaluate across the venue;
  • one login and streaming session across supported markets; and
  • a concentrated liveness boundary and shared capacity across those markets.
The result is not “trustless matching.” It is a narrower, inspectable trust model: trust the measured matcher for ordering and price policy; verify Solana and the proof system for custody and settlement validity. Each market still has its own governed mint pair, oracle policy, and order book, and one proof batch can contain only one market. A finalized governance or signer mismatch pauses new trading venue-wide while cancellation and settlement reconciliation continue. Authenticated-oracle staleness or replay failure pauses only markets bound to the affected feed; each /instruments entry reports that market-local readiness. Oracle, governance, and drain reasons remain independent, so recovery of one subsystem cannot accidentally clear another. See Multi-Market Venue.

​
What an integrator should verify

Before sending private order intent, verify the fresh quote, replay the measured event log, approve the measured release, bind the complete signer set, and compare it with finalized on-chain configuration. Pause new trading when that evidence or the on-chain key view becomes stale. See Privacy & Attestation for the verification chain and Settlement for the proof boundary.