SilentEdge

Non-custodial Solana trading vaults where your strategy thresholds are evaluated inside Arcium's MPC network, so the platform operator never sees them.

Status — read this first

Working on devnet. Not audited. Not on mainnet. Do not use with real funds.

The program's upgrade authority is a single hot key — a plain system-owned keypair, not a multisig. Every custody statement below describes the bytecode running today, not a promise about tomorrow's: whoever holds that key can replace withdraw. Moving it to a multisig is the one blocker on mainnet. scripts/check-upgrade-authority.mjs refuses a deploy while the authority is a plain keypair, and exits 1 today.

What	Grade	Evidence
Browser-encrypted strategy → MPC evaluation on live Pyth SOL/USD → BLS-attested callback → withdraw	Verified end to end against the live devnet cluster	tests/e2e-devnet.ts
The swap that spends an authorization (devnet has no routable liquidity)	Verified on a surfpool mainnet fork with a live Jupiter route	tests/swap-execution.ts
Cluster pinning (T-37), the most custody-critical check in the program	CODED, not ENFORCED — the derivation has Rust unit tests, the call site has no runtime detector. A previous test claimed to catch its removal; a mutation run proved it did not.	SECURITY.md
Program upgrade authority (T-3)	UNVERIFIED	SECURITY.md

The last full pass found eight real issues: seven fixed, one not — T-37 above. Two of the fixed ones would have cost a user money: a cooldown that disabled the stop-loss, and a strategy replacement that left the old strategy trading. Both now have detectors confirmed by deleting the fix and watching the test fail. Enforced, merely coded, and claimed-but-absent are graded threat by threat in SECURITY.md.

What it does

1. Connect a Solana wallet and create a vault — a PDA controlled by the program, not by us.
2. Build a strategy visually, e.g. buy SOL below $150, sell above $180, stop at $120.
3. The strategy is encrypted in your browser before it leaves it, and stored on-chain as confidential state.
4. Arcium's MPC network evaluates it against a Pyth price and outputs only BUY, SELL or HOLD. Inputs are secret-shared across nodes; under its dishonest-majority model your parameters stay private as long as at least one node is honest.
5. The result returns in a callback carrying a BLS threshold signature from the whole cluster, verified on-chain in our program. The operator cannot forge one; a compromised backend can only refuse to schedule work — a liveness failure, never a safety one. Valid results execute via Jupiter.
6. You withdraw to your own wallet whenever you want. Funds leave a vault by exactly two paths: a swap between allowlisted mints where both sides stay inside the same vault, or a withdrawal signed by the owner. No instruction in the deployed program accepts an operator authority — a property of the instruction set, not a runtime check.

What this project does not claim

• The program can still be changed. No instruction lets the operator withdraw your funds, but until the upgrade authority is a timelocked multisig, the operator can replace the instructions.
• Your strategy does not stay hidden forever. Every executed trade is public, and an observer correlating prices against your trades narrows your thresholds over time. That is inherent to acting on a public chain; no cryptography fixes it.
• We hold the Arcium MXE authority. Using it would halt every bot visibly on-chain and cannot forge a trade, but it would let us decrypt previously stored strategies. There is currently no way to burn or timelock it.
• Trade intent is briefly public before execution. Front-running is possible.
• Not for HFT. MPC evaluation takes seconds. Suited to threshold strategies over hours and days; unsuited to arbitrage, scalping, or stop-losses that must fill at a price.

Running it locally

Fast path — the web app against the already-deployed devnet program. Copy .env.example to .env and fill it in, then:

pnpm install
pnpm --filter @silentedge/web dev        # http://localhost:3000

Full path — build, deploy and test.

pnpm surfpool                            # terminal 1: surfpool, forked from devnet
anchor build                             # terminal 2
anchor deploy --provider.cluster http://127.0.0.1:8899
pnpm test:local                          # vault + authorization, against the fork
pnpm test:devnet                         # end-to-end, against the live devnet cluster
pnpm test:fork                           # swap execution, mainnet fork + live Jupiter

Forking devnet means the real wSOL and USDC mints exist, so the production allowlist is genuinely exercised (docs/testing.md).

Devnet SOL. A deploy needs roughly 5.3 SOL of headroom for the temporary buffer, refunded afterwards. solana airdrop is rate-limited to the point of being unavailable, so top the deploy wallet up at https://faucet.solana.com or from another devnet wallet first.

Never put a funded keypair path in .env. The backend holds no key that can move funds.

Repository layout

Path	Contents
programs/vault/	Anchor program — custody, limits, intents, swap CPI
encrypted-ixs/	Arcis circuits (Arcium workspace convention)
programs/hello_arcium/	Disposable proof the Arcium pipeline works; its strategy circuits are retired
apps/web/	Next.js — dashboard, strategy studio, client-side encryption
apps/api/	Untrusted service — scheduling, indexing, RPC relay
packages/	sdk strategy encryption (RescueCipher + x25519), types model and validation, config mints and allowlists
tests/	Local-fork, live-devnet and mainnet-fork suites
scripts/	Upgrade-authority check, circuit registration, fork seeding

Verified toolchain: Rust 1.97.1, Solana CLI 3.1.11 (Agave), Anchor 1.1.2, Arcium 0.14.1, Surfpool 1.5.0, Node 24.7.0, pnpm 11.21.0.

Documentation

Document	What it is
SECURITY.md	Every threat graded ENFORCED / CODED / UNVERIFIED, plus disclosure policy. Read this before trusting any claim above.
ARCHITECTURE.md	The design, and the data, money and strategy flows
docs/privacy.md	What the engine hides, and what leaks anyway
docs/arcium.md	How the MPC pipeline and persistent encrypted state work
docs/testing.md	Test environments, coverage, and a false-pass trap
docs/research.md	Verified findings from official docs, with sources

Also: FEES.md (no protocol fee — a fee would be a third way for value to leave a vault), docs/oracle.md (why Pyth triggers and Jupiter quotes), and docs/magicblock-evaluation.md (why an Ephemeral Rollup cannot sit on this critical path).

License

Apache-2.0. Permissive, with an explicit patent grant — the licence Solana and Anchor themselves use.