Prism

How we reviewed this: we installed it in Docker with Bun and ran its full test suite offline. We deliberately ran nothing that touches a chain, a wallet or a key: no setup, no registration, no enter, exit, rebalance, swap or harvest. Everything about on-chain behaviour here comes from reading the code.

What it does

Prism, by irfndi, watches Meteora DLMM pools on Solana and decides whether to hold, rebalance, exit or enter. It is a large codebase — one engine file runs to 15,000 lines — and 2,167 tests pass offline in about two minutes.

Jev occupies two small files. For each entry candidate, one request asks four questions: which deposit distribution fits this regime (Choice), is the volume toxic flow, is a return to range likely enough to hold, and is the pool in systemic stress.

Jev cannot open or close a position. Its only output with any effect halves a position size, and that path is guarded by paperTrading === true — in live mode it is unreachable, and it is off by default. The rest is logged, and written to a vector memory that a separate LLM layer may later read as context. That’s an indirect channel, and worth knowing, but the hard rule holds: the typed decisions don’t move money.

The state it sends is clean: pool address and pair, TVL, volume, fees, a few computed ratios and the active bin. No wallet, no position size, no balances. Requests are rate-limited behind a circuit breaker.

The trust boundary we’d want changed

Fee-sharing configuration — the platform fee rate and the wallet it is paid to — is fetched from a remote endpoint the project controls. The parser takes both values as they arrive, and the downstream check accepts any rate up to and including 100%. The resulting transfer is appended to a transaction signed by the operator’s wallet. There is no local ceiling, no allowlist of destinations and no confirmation; the one guard is a check that the destination isn’t the wallet itself.

Nothing suggests bad intent — the published tiers are ordinary percentages, and this is how a hosted service funds itself. The problem is structural: if that endpoint is ever wrong, compromised or impersonated, the local software will sign what it is told. A hard cap in code and a pinned destination would remove the whole class of risk.

Because this project handles live funds and has no private reporting channel, we asked the author to enable one rather than publish details. We’ll follow up when there’s a way to send them.

Two related notes for anyone running it: the shipped token and deployer blacklists are empty files, and the README is candid that every external screener is advisory and fails open. And the default backtester runs on synthetic data — its own banner says the results are a lower-bound stress test that assumes no token price depreciation.

Verdict

The discipline here is real: the model advises, code decides, and the one place a typed answer can change a number is fenced behind paper trading. The test suite is the largest we’ve run in this ecosystem and it is green offline.

If you’re studying how to build with a model near money, read this. If you’re thinking of running it with funds, read the revenue-share path first and decide whether you want a remote endpoint to have that much say.

For other finance projects, see jev-trader in our roundup.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 20, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.