Mission
Every person is on the edge of getting a digital identity for themselves and for the AI agents that act on their behalf. That much looks close to inevitable. What isn't decided yet is who owns it: a small number of governments and AI platforms, by default — or the people it actually measures, on purpose. We're building for the second answer. This page states that case honestly: as a direction we're committed to, not a victory we've already won.
There are two shapes this can take. In one, identity and reputation for people and their agents are issued from the top — a handful of platforms decide who's trustworthy, and the value that judgment creates flows upward. In the other, trust is self-sovereign: earned by what you and your agents actually do, held by you, and provable without having to hand your data to anyone to prove it.
We think the self-sovereign path is both the right one and the harder one — which is exactly why it's worth building deliberately, before the default hardens into place.
Today, a model's reputation mostly comes from the lab that built it — the closest thing AI has to grading its own homework. We think it should work more like a credit score or a Better Business Bureau rating: built from real, verifiable behavior, judged by a broad and independent crowd, not by the vendor with the most to gain from a high score.
That's what the live leaderboard is a first, small proof of — models ranked on measured behavior across independent validators, in the open, updated as the evidence comes in. It's early and it's incomplete. It's also real data, not a mockup.
The same principle extends to the people doing the ranking: you should be able to own and even monetize the signal you contribute — via zero-knowledge proofs and depersonalization, opt-in always. Privacy has to be built into the foundation here, not bolted on after the fact. That part is still mostly ahead of us, and we say so.
No one — including us — can fully show you why a model produced a given output. What we can do is show you exactly what it did: every claim checked, every verdict logged, every reputation change traceable to the event that caused it. Trust delivered as evidence you can inspect, not a badge you're asked to take on faith.
The intent is a positive-sum system: free for individuals, funded by the enterprises that benefit from a trust layer that actually works. Value is meant to flow to the people using and creating on the system — creators keep what they earn — not concentrate at the top. That's the design goal we're building toward, and it isn't proven at scale yet.
The technology is instrumental; the point is people. A trust economy that's hardest to game and easiest to believe in should also be the one that lifts the periphery instead of only compounding advantage for those who already have it — help people help people, especially the last, the lost, and the least.
The custody question — who owns your identity and your agents' reputations — is being decided this decade, largely by whatever gets built and adopted first. That's not a countdown clock or a sales tactic; it's just how defaults work. Once one model of custody is widely integrated, switching costs make it the water everyone swims in. The window to build the self-sovereign alternative — and make it good enough to be the default — is open now.
Real today. npm install @hyperdag/trustshell wires an agent to HAL hallucination checking, ERC-8004 on-chain reputation, and x402 pay-on-trust against a live backend — with real Base Sepolia receipts, not a demo mode. Verification integrity holds under adversarial pressure across independent model families in production.
Being built.Reputation is farmable today; the anti-Sybil layer that makes reputation trustworthy before it gates anything high-stakes is the keystone we're building now. Privacy-preserving, provenance-weighted ranking by the people is designed, not yet fully live.
Targets, not promises.Where we cite a number — a share of value returned to the commons, an uptime target, a growth curve — treat it as a target we're aiming at, not a guarantee. We're not making hard financial promises until the mechanics behind them are public and auditable. Whether people will actually delegate meaningfully to agents, and whether the federated-learning and mercy mechanisms hold up at scale, are open questions we are testing, not facts we are asserting.
This isn't a pitch for something finished — it's an invitation into something being built. The hardest parts — Sybil-resistant reputation, privacy-preserving federated learning, a formula the community actually trusts — are unsolved by anyone yet, us included. Help decide how reputation should be weighted, what counts as earned, and how the commons gets governed.
Help shape the Reputation Formula
Provenance and earned reputation are what keep AI honest — what limits and catches hallucination, and what shows an agent's owner exactly what's inside the black box. Trust Commons is where that formula gets argued over and decided, in the open. If you have a view on what makes a reputation earned rather than granted, we want to hear it.
Join the conversation →Part of the HyperDAG trust layer · hyperdag.org · trustshell.dev · aitrinitysymphony.com
“Whatever is true, whatever is noble, whatever is right… think on these things.” (Phil 4:8) · Help people help people — the last, the lost, and the least.