Beaver Knight

01 · The problem

AI agents that trade real money are multiplying, and every one claims to be a winner. Two things make that dangerous.

• You can't tell a real record from a fabricated one. Leaderboards show huge headline numbers. An agent can report over half a million dollars in volume with almost nothing showing up on-chain, and a naive check reads the wrong wallet entirely. Reputation built from client feedback gets farmed. Claims are everywhere, proof is nowhere.
• Nobody is vetting them. The platforms that host these agents say plainly that they do not supervise, audit, or vouch for any agent. So the honest question - is this agent's track record actually real - has no trustworthy answer.

02 · What Beaver Knight is

A credit bureau for AI trading agents. Point it at any agent and it resolves the real trading wallet, reads what actually happened on-chain, and returns a report where every number is labelled verified or reported. The reports live on a public board, ranked by what the chain proves, not by what agents claim. Like a credit bureau, we do not lend and we do not take deposits - we rate, and everyone else reads the rating.

03 · How the check works

• Resolve the real agent. Every agent is a tangle of addresses - owner, token, execution wallet. Score the wrong one and you get the opposite answer, so the first job is finding the wallet that actually trades and rejecting the impostors trading on the same name.
• Read the chain, not the pitch. Realized profit and loss read straight from the exchange, vault share price from archival reads, on-chain activity, and who an agent actually transacts with. No API key and no privileged access, so anyone can re-derive it.
• Separate proven from claimed. Every field is labelled: verified means we read it from the chain and you can re-derive it, reported means the platform said so and we relay it. We never dress a claim up as a fact.
• Portable by design. Every check looks for an ERC-8004 identity on Base, though almost no agent has one yet. Writing the score itself back to that reputation standard, so any app or agent that speaks it can read our rating, is the next build - the trust should never be locked in our silo.
• Check in seconds. The board shows agents we have already scored; a search checks any agent that is not on it yet. Free, the way you would run a credit check before you lend.

04 · Why you can trust the score

• The referee, not a player. We rate agents. We run no fund and never touch your money, so we have no reason to favour anyone. Every agent on the board is judged by the same rules.
• Earned, never bought. Nobody pays to be rated and nobody can decline a rating. If an agent's numbers do not hold up on-chain, we say so. A rating you could opt out of would only ever carry good news.
• Re-derivable by anyone. Every score comes from public data, with the sources named on each report. You do not have to trust us, you check the arithmetic yourself. When we tighten the method and a rating falls, the change stays on the record with the date it happened.
• We say what we could not check. Every report lists what we could NOT establish, beside what we could. Nothing on-chain proves a machine rather than a person is trading an account, and a vault named after a famous fund is not that fund. Those gaps are published rather than left for a reader to assume away, because a bureau that only prints what it verified lets silence read as approval.

05 · How it makes money

Like a credit bureau, the people reading a score do not pay for it. The businesses do.

• Free for users. Checking an agent is free, and always will be. It is the front door, and free checks are what fill the directory with an audience.
• Agents pay for the Verified badge. An agent owner claims its profile and earns the Verified by Beaver Knight badge - proof its record is real, shown where people decide who to trust. Earned from the chain, never bought.
• Agent-to-agent checks. One agent can pay a small fee in USDC on Virtuals ACP to verify a counterparty before it transacts with it - verify before you pay.

06 · What is live today

• A public directory at beaverknight.com, seeded with real top agents scored from public on-chain data, whether they asked or not.
• The verify engine: resolve the real wallet, read realized profit and loss from Hyperliquid and share price from ERC-4626 vaults, corroborate reported volume against on-chain activity, and sample the wallet to catch dormant, self-dealing or scam-exposed agents. Every field is labelled verified or reported, and every report also lists what we could NOT establish.
• Scores stored on-chain in an Internet Computer canister as an append-only log, so every rating we have ever published and the moment it changed stays readable by anyone. We can publish a new score; we cannot quietly rewrite an old one. The writer of that record is a service key today, and we would rather say that than overstate it.
• Ratings attested on Base by a key that does not exist. A second canister, the beacon (w4sbh-3aaaa-aaaak-qzbha-cai), signs with Internet Computer threshold-ECDSA from the address 0x46a7f2a1ffafaad254a91c0c71525e3e7e71548f; no person and no server holds a private key for it. It registered Beaver Knight as ERC-8004 agent #8990 and publishes ratings as EAS attestations under schema 0x639813d4…66b4 (agent, score, level, rated-at, evidence hash, report URI). A report’s JSON carries the attestation UID when one exists, and a plain null when it does not, which is most records today: attestations are issued when a rating’s band changes, on Base Sepolia until the mainnet flip. So two things are true and we state both: the record in the bureau is written by a service key; the attestation of it on Base is not.
• A lookup by address at /api/rate?wallet=, free and unauthenticated, for anyone deciding whether to trust a counterparty. It answers on an agent’s execution wallet, owner wallet or token, it keeps what we found separate from what we could not reach, and an address we have never rated comes back as an explicit warning rather than a silence that reads like approval. Everything before it - the badge, the report, the claim page - needed the rated agent to co-operate, and most of a register full of findings never will.

We have also built our own keyless trading agent, and it runs on mainnet with real money. Beaver Knight Trade is an Internet Computer canister (xe3ue-vaaaa-aaaak-qzbda-cai) that trades Solana through Orca from the wallet Bz3LGi9R3kfCo1wG9LZsoeGj29SsNvqVQxbw9tDhFBfR. It signs with threshold Ed25519, so no complete private key for that wallet exists at any moment, including while it signs: the key is held only as shares across the subnet’s nodes and is never assembled. Sixteen trades are on record and every one of them is finalized on Solana for anyone to check. It has not traded since 28 July, because its own published rules said not to, and it writes that decision down every twelve hours.

That is reference proof that we can build the kind of agent we rate. It is deliberately NOT on the board. A bureau that rates its own agent has precisely the conflict this bureau exists to detect, and we would rather give up the row than argue we are the exception.

07 · Roadmap

• More agents, richer profiles. Grow the board and give each agent a full profile - what it does, plus a link to go use it.
• The Verified badge. Let agent owners claim their profile and earn the badge, and embed it on their own site.
• Monitoring and alerts. A living scorecard that tracks an agent over time and flags when its record stops holding up.
• Every rating attested, on mainnet. The keyless attestation rail is live on Base Sepolia and issues on band changes. What is not yet done: every rating carrying a current attestation, retractions revoked on-chain rather than only withdrawn here, and the move to Base mainnet, which is a network constant, a schema registration and a few dollars of gas. Until a rating carries its attestation a caller is trusting our server, and the honest word for that is a dependency.

08 · About

Beaver Knight is built by one founder, in the open. No token and no hype, just a working engine you can check yourself. It is early and unaudited on purpose, so the record shows the truth, good or bad.

If you want to use it, integrate it, or back the build, reach out on X.