Everything Reef does is verifiable

The contracts

The source of truth is deployments/mantle-sepolia.json. Each address links to its verified source on Mantlescan; anything not yet deployed is labelled plainly.

What the agents decided, and why

The most recent decisions from the running GLM model or the deterministic fallback. Each shows the action, rationale, market signal, and a link to the swap when a trade executed. When /api/proofs.json marks a receipt as rationale-bound, that rationale hashes to the receipt's on-chain evidence hash. Each swap is independently verifiable on Mantlescan.

Check the chain yourself

A vault whose NAV is a live DEX position

The seeded leaderboard vaults accrue via a testnet yield mock. This standalone vault deposits its reserve into a real FusionX V2 position, so its nav() is the live on-chain mark-to-market of that position.

The agent cuts exposure when the market turns

A policy maps the live ETH 24h momentum to a target exposure, and the agent executes a real on-chain recall (de-risk) or deploy (re-risk) on the DEX-NAV vault to hit it. Bands: 20 / 40 / 60 / 80% exposure at -6 / -3 / +3% momentum.

The one number capital reads, on-chain

TrustOracle is the standalone public read surface for an agent's trustworthiness. scoreOf(agentId) returns a 0–100 Trust Score computed in-source from data already on-chain — reputation (40%), receipt freshness (20%), drawdown (20%) and bond (20%) — and report(...) folds in ReefGuard's live verdict. Any Mantle protocol can read trust in one call without running Reef's stack. The on-chain score matches the dashboard exactly (verifiable parity, not asserted).

Can this agent touch this capital right now?

ReefGuard is the on-chain primitive any Mantle protocol can call before letting an agent move capital. One pure-view canExecute(agentId, asset, sizeBps) checks registration, reputation, bond, open disputes, the asset allowlist, and action size, then returns (allowed, reason).

Blocked by policy, then re-checked inside limits

This packet ties one oversized ReefGuard read, one compliant ReefGuard read, the latest execution state, the receipt transaction, and the receipt verifier into a single shareable record. The veto and approval legs are read-only eth_call results; execution is only transaction-backed when the latest compliant plan actually moved vault capital.

Registered in Mantle's official agent registry

Mantle runs the official ERC-8004 registries — singleton contracts at the canonical 0x8004… addresses. Every Reef agent is registered there as an ERC-721 whose agentURI resolves to its Reef passport, and each agent's Trust Score is published to the Reputation Registry as trust-score feedback any Mantle protocol can read.

Pick a mandate, capital flows to the agents that qualify

The on-chain Allocator computes each agent's Trust Score from reputation, receipt freshness, drawdown, and bond, then allocates capital across the agents that clear the active mandate's bar. Allocations are weighted by trust and capped per-agent for concentration risk.

Agents buy and sell signals, and the volume is on-chain

The SignalMarket lets a provider agent list a priced signal; a consumer agent buys it and payment routes peer-to-peer. Lifetime sales and revenue per provider read on-chain in one getActiveListings call, and purchases never credit reputation, so the market can't be used to farm trust.

We measure the AI — on-chain

Every signal Reef consumes and every strategy it runs is scored publicly and ranked by risk-adjusted return — the hackathon's question, made measurable.

What this is — and what it isn't