A language model is sharp at reading intent and choosing the next move, and weak as a safety boundary. Lyra keeps it advisory — it interprets your goal, picks tools, and surfaces opportunities, but never holds the final word over your funds.
Before value moves, the action clears a deterministic policy written in Move and enforced on Sui: budget and per-transaction caps, coin and recipient allowlists, slippage limits, an autonomy tier. No model, no network in the loop — a breach reverts on-chain.
Every proposed PTB is dry-run against a live Sui fullnode first. A failing effect aborts the action with a decoded reason, so a transaction that would fail never reaches the chain or burns a unit of gas.
The approval floor lives under the session mode. A material-risk action asks for your sign-off even when autonomy is wide open, and is refused outright under strict. The controls live in Move — not in a sentence the model could talk its way past.
Cleared actions spend from your on-chain vault, wait for the effects, and return a decision record — the policy verdict, the dry-run, the digest, and a Move ActionReceipt. Memory and receipts persist on Walrus, reconstructable after the fact.
Balances and transfers, swaps routed through Cetus, FlowX, Bluefin and DeepBook, lending on NAVI, Scallop and Suilend, and yield discovery via DefiLlama — driven from the CLI, Telegram, or the web console.
Chat it, run it, or build on it. The policy engine keeps the boundary.
