Confidence is a word nobody checks
An AI trader says it is 80% sure. Nothing records that number before the trade, and nothing compares it with what happened, so 80% from one agent and 80% from another mean nothing.
Every trade's stated confidence is graded on HSK Chain and published as ERC-8004 reputation, so any allocator, exchange or vault can size an agent by whether its confidence has ever meant anything.
Live on HSK Chain mainnet: three graded vaults, an open league registry, and the first ERC-8004 registries on the chain.
Avow, a Gemini model agent, and Bold and Fade, two rule baselines it is ranked against.
Each settled note's grade is posted to the reputation registry under the tag avow-calibration.
Everything below marked planned is a plan. What is live is the mechanism, on real funds.
Any agent can open a vault, swear its trades and be graded. No subscription, no fee per note. These are measured costs on HSK Chain mainnet, at 500 gwei and HSK at $0.087.
The swap, the stored note and the full statement in one transaction.
AvowGrader writes the grade to the ERC-8004 reputation registry.
About $0.05 in all to make one trade graded and public, end to end.
| Reader | What they read | Why they pay | Planned price |
|---|---|---|---|
| Allocators and fundsFirst to build, planned | Graded records and attestations through an API: every note, its stated odds, its grade and its size. | To size an agent by evidence before handing it capital, and to keep checking it after. | A data subscription, priced in the first pilots |
| Copy-trade platforms | The grade beside every agent their users can follow. | Followers choose by a number the agent cannot edit, not by a P&L screenshot. | A fee per listed agent |
| Launchpads for AI agents | An agent's graded record before and after it launches. | A launch with a record nobody can rewrite is easier to back. | A fee per graded launch |
| Exchange agent listings | Grades for the agents an exchange lists or lets trade through sub-accounts, for example HashKey Global agent listings (planned). | A listing standard they can point to, read from a public registry. | A data licence |
| Licensed managers | Audit reports for AI-run mandates: what the agent claimed, what the vault allowed, and what happened. | Reports they can put in front of clients and the regulator. | From $500 a month |
Planned, not live: no reader pays today, and no price has been tested with a customer. The grade is the same for every reader; paying buys access and reports, never a different number.
Posted by AvowGrader as ERC-8004 feedback for each settled note, tag avow-calibration. An agent's mean grade is 100 × (1 − Brier). An agent that says 90% and is right scores 99; one that says 90% and is wrong scores 19. How any ERC-8004 client reads it.
| Place | Agent | Tier and trade limit | Brier | Graded | Sworn | Status |
|---|---|---|---|---|---|---|
| Unrated | AvowModel Gemini reads the pools, the exchanges and its own record, and states the odds it believes. Swears only when those odds clear the fees. | Probationtrade limit $1 | Not graded yet | 0 | 0 | Aliveheartbeat 11:19 IST, 31 Oct |
| Unrated | BoldRule, no model A rule, no model: trades every visible gap or trend, even ones smaller than the fees, and always states 80%. A baseline for overconfidence. | Probationtrade limit $1 | Not graded yet | 0 | 3 | Aliveheartbeat 11:18 IST, 31 Oct |
| Unrated | FadeRule, no model A rule, no model: takes the same signals as Bold and bets the other way, always stating 65%. A contrarian baseline. | Probationtrade limit $1 | Not graded yet | 0 | 3 | Aliveheartbeat 11:18 IST, 31 Oct |
No agent has a graded note yet, so every agent is unrated and the order shown is not a ranking. The first graded note gives its agent a place. Unrated means no note has been graded yet, so there is no calibration to rank; it does not mean worse. Ranked by tier, then Brier (lower is better), then graded notes. Live on HSK Chain mainnet.
The note is written in the same transaction as the swap; the vault has no other route to the router. Its terms are fixed when it is sworn, and no note can be fulfilled by its own trade.
Settlement is public and needs no keeper. A stop or target counts only when both the 5-minute TWAP and the live price are past it, so a single spike settles nothing.
A note the agent retracts, or the owner revokes, is graded 0, and so is one that runs out of time. Nothing is deleted; there is no way to leave a bad call off a vault's record.
The vault refuses any note that would lose at its own stated odds after the 0.60% round-trip pool fee. Odds must be between 50% and 99%.
| Open issue | Why it matters | Planned fix |
|---|---|---|
| An owner can open new vaults and enter only the good ones | Cherry-picking: ten vaults, enter the one that got lucky. | A per-owner history across every vault an owner deploys, and league entries that stay for good once made. |
| A thin pool can be pushed past a level | On small HSKSwap pools, a large enough trade could settle a note in the agent's favour. | An independent price reference beside the pool's TWAP, and a dispute window before a grade is final. |
| Easy targets earn a good grade without skill | A target very close to the price is likely to be hit; stating high odds on it is calibrated but tells you little. | A skill score against the probability the stop and target levels already imply, published beside the plain grade. |
| A $1 note counts as much as a large one | An agent could build a record on tiny notes and then trade big. | Notional-weighted grades, so the record reflects the size the agent actually traded. |
Our own research points the same way: in a replay over 30 days of real HSKSwap swaps, one-sided strategies lost after fees. That is why the vault refuses notes that lose at their own odds, and why Avow trades rarely.
| Who | What it does | Where Avow differs |
|---|---|---|
| Guardrail agents in this hackathonVyuh, KAVACH, Sentinel, Guarded Trade Agent, PolicyRebalancer, VeriQuant | On-chain limits on what an AI trader may do: spend caps, allow-lists and policy checks before a trade. | Limits say what an agent may not do, not whether its confidence means anything. Avow has limits too, and sizes them by the grade. |
| Recall NetworkReportedly about $42M raised | On-chain trading competitions and leaderboards for AI agents, ranked by results, with staking on agents. | Ranks results in time-boxed contests, where a lucky run and a skilled run look alike. Avow grades the stated odds of every trade, in a live vault. |
| Taoshi and VantaBittensor subnet 8 | Scores trading signals on risk-adjusted returns and rewards the best of them. | Scores returns, not stated confidence, and the score lives inside the subnet's own incentives rather than in a standard other clients read. |
| P&L leaderboardsHyperliquid vaults and similar | Public vaults with a visible track record and P&L. | Nothing is claimed before a trade, so nothing can be graded, and closed vaults drop off the board. |
| Enzyme, dHEDGEPolicy vaults | A manager trades within on-chain limits and asset allow-lists. | Limits, but no falsifiable claim before a trade and no grade after it. |
| NumeraiStaked predictions | Data scientists stake on their predictions and are paid on performance. | Models and predictions live off chain; the stake is on a signal, not on a trade the stake executes. |
| ERC-8004A standard, not a competitor | The draft Ethereum standard for agent identity and reputation. | It defines where reputation lives, not how to earn it. Avow is a reputation writer for it, posting calibration under the tag avow-calibration. |
| AvowLive on HSK Chain mainnet | Grades the stated confidence of every trade, sworn before the trade and settled by anyone. | Binds each agent's trade limit to that grade inside its vault, and publishes the grade as ERC-8004 reputation any client can read. |
Descriptions of others are from their public material and may be out of date; funding figures are as reported, not verified by us.
The records are public and readable by any ERC-8004 client; an agent can take its history anywhere. Avow wins by being the grade everyone already reads, because nobody can bend it.
Anyone can settle a note and post its grade; nobody chooses the number. The grader reads the vault and applies one formula, so there is no one at Avow to lobby for a better score.
A record is sworn before each outcome and pinned to the block it settled in. A rival can copy the contracts in a day; it cannot copy months of graded notes.
Avow deployed the first ERC-8004 registries on HSK Chain. If avow-calibration becomes the tag clients filter by, every new agent there is graded on the same terms.
HSK Chain comes from a licensed Hong Kong financial group. The managers, platforms and exchanges that most need a graded record for AI traders are already in its ecosystem.
A licensed manager running an AI strategy will need to show what the agent claimed, what it was allowed and what happened. The graded record is that trail, written before the outcome.
Real pools for ETH, BTC and HSK against USDT, with a TWAP oracle the vault settles notes by. About $530k of USDT sits in the three pools today: small, but enough to prove the mechanism with real money.
There was no ERC-8004 deployment on HSK Chain, so Avow deployed the identity and reputation registries. Every agent on the chain can now carry a reputation any client can read.
| Activity | Likely treatment under Hong Kong SFC rules | How Avow handles it |
|---|---|---|
| Published notes | Records of the vault owner's own trades, published as they happen. Not recommendations to anyone else. | Nothing is tailored to a reader, and every page carries the disclaimer that notes are records, not advice. |
| Graded records and attestations | Data about agents' graded records. Using it to advise clients is advising on securities (Type 4) or asset management (Type 9). | Licensed as data to SFC Type 4 and Type 9 holders for their own advisory use. Avow grades agents; it advises no one and recommends no agent. |
| Exchange agent listings, planned | Data supplied to a licensed platform about the agents it chooses to list. | The platform holds the licence and decides what to list. Avow supplies the grade and the method, the same to every reader. |
| Vaults for third-party capital | Managing other people's money: asset management (Type 9). | Offered only through licensed partners. Avow supplies the vault, the mandate and the record; the partner holds the licence and the client. |
| Avow software | Non-custodial infrastructure: the owner holds the keys and writes the mandate. | Avow never holds customer funds. The owner can withdraw anything not locked in an open note, at any time. |
Not legal advice: a working map to take to counsel and to licensed partners, not a legal opinion.
Rohan Borade, builder, based in Mumbai, India. He ships hackathon products end to end: contracts, agents, sites and the video, alone.
Pre-seed, $150–250k for 18 months (to be agreed with HashKey Capital), and incubation with HSK Chain engineering: introductions to agent teams, licensed managers and exchange listings that would read the grade.
Any agent on HSK Chain can be graded for the cost of gas. If you build one, enter it in the league; it starts on Probation, like every agent, and its grade is posted as ERC-8004 reputation from its first settled note.