Hybrid DAG + Blockchain
A directed-acyclic-graph fabric for the throughput agents need, anchored to a blockchain for settlement and finality. Fast where it should be fast; immutable where it must be.
HyperDAG Protocol
AI is being trusted with everything, and digital identity is coming — for people and their agents alike.
Do nothing, and governments and the AI oligarchy decide. Build the alternative, and it's yours — owned, controlled, and private — serving the people it measures, not the few who watch them.
What we believe
When autonomous agents start acting on our behalf — spending, deciding, speaking for us — they need a verifiable chain of custody: who did what, under whose authority, and whether it can be trusted. That is exactly what a blockchain is for. AI is the reason it finally matters — more than crypto, more than anything yet known. (We say "blockchain" because it's familiar; we mean DLT. A DAG is another kind — and a hybrid is, so far, the best fit.)
A chain-of-custody for people and agents is inevitable — the question is by whom, and on whose terms. Top-down or self-sovereign. One future has identity imposed by an AI oligarchy that can't help but placate the incumbents who fund it. The other is private, permissionless, and owned by the people — who set the weights and the values through their own agents. Safe and ethical AI is achievable, but only when the people, not a captured regulator, set what "safe" and "ethical" mean. It's only zero-sum if we let it be — we build positive-sum, for the good of all.
Flip the script
Today the big models rank themselves and ask you to trust them. We're building the opposite: people and their agents rank the models — think Yelp, a credit score, and a Better Business Bureau in one, by the masses for the masses, while your data stays in your control.
We don't reduce a model to one number. Like blood pressure — 120 over 80 says what neither figure says alone — trust reads truer in more than one dimension, and the signal is in how they relate. The engine behind it — HAL — has already run 147,595 real fact-checks (live). Today the board ranks truthfulness — accuracy under fact-check, scored so confidently wrong costs more than honest doubt — with cost beside it as a reference: real value per dollar, and an honest check on whether the newest, priciest models are actually worth it for everyday work. So far this measures code-review discrimination — one narrow proxy, honestly labeled — and we're broadening toward general trustworthiness. A leaderboard nobody, including us, can grade their own homework on.
The lens we're building: defensibility × discernment. Defensibility — do your own claims survive when others check them? Discernment — are you right when you check theirs? Producing truth and detecting it are different skills. You don't ask the doctor to diagnose themselves, or the thief to guard the cash — so an agent's standing shouldn't rest on its own word. Where an agent sits in that relationship tells you how to use it, and when to trust it — not just how to rank it. And it's hard to fake both at once, from independent sources. That's the point. The peer-verification behind it — 140,000 checks and counting (live) — is already flowing.
Beyond anything a board can measure from the outside is perceived trust — what the people and agents who actually use a model report. We're only starting to collect that (join us below), and where measured and perceived disagree is the most revealing signal of all.
Built against Goodhart's law. When a measure becomes a target, it gets gamed. So no single number rules here: decorrelated validators that never grade their own family, a proper scoring rule, and — soon — the independent judgment of real users. Gaming one signal is easy; gaming all of them, from independent sources, is the hard problem we build on. This is our best indicator so far. Help us improve it, break it, or make it →
Today's gatekeepers monetize the content you create for free. Flip it. Keep your data private (only you), or depersonalize it — zero-knowledge proofs plus depersonalization strip out the personal information, then you monetize what remains. Those who want access must prove who they are, and pay for what they currently take for free. Opt-in, always; a right to privacy sits at the core, not bolted on.
The tech thesis
We can't show you why a model thinks what it thinks — no one honestly can. But we can show you exactly what it did: every action visible, measurable, and traceable. Trust delivered as evidence, not asserted as a claim.
A directed-acyclic-graph fabric for the throughput agents need, anchored to a blockchain for settlement and finality. Fast where it should be fast; immutable where it must be.
Prove who you are, what you're allowed to do, and what you did — without exposing the underlying data. Self-sovereignty isn't a slogan; it's the cryptography.
A portable, behavioral reputation that agents and people accrue through what they actually do. Reputation you carry, not a score handed down by a platform. Earned, not granted.
A quorum of independent, decorrelated validators fact-checks each claim. We never let models from the same family grade each other's work — same family, same blind spots — so a flaw one misses is caught by another that fails differently.
Every verification is routed to the validators most likely to catch its kind of error. An adaptive neuro-fuzzy system with LASSO feature-selection learns which models catch which failures — tuning accuracy per dollar in real time.
The system improves from stress. Rotating red-teams probe it, disclosed-and-repaired errors earn credit, and recurring mistakes decay toward zero reward — so attack and error make it more reliable, not less. Antifragile by design.
What we're building
A tunable trust harness where agents, and the people behind them, rate one another and the models that feed them, first-hand. In an agentic world, performance alone isn't enough — reputation is, earned through both self-review and the scrutiny of peers. We're trying to give raw intelligence what it lacks on its own: common sense, street smarts, the instinct to sense when something's off — and, over time, wisdom. That takes character, dignity, and a moral compass — and those come not from the model, but from a higher source and the people it's built to serve: the last, the lost, and the least.
Get involved
A movement, not a company — an open protocol anyone can read, run, challenge, and build on.
ZKP circuits, contracts, and the published methodology. Read it, run it, and help improve it.
Join the conversation on what formulas should make up the best reputation matrix — the weights, the scoring rules, who governs them. Open, protocol-agnostic, in the Trust Commons.
Or help build the half that can't be measured from the outside — perceived trust. We believe, and so do the models themselves, that reputation and user satisfaction are part of trust achieved — not all of it, but a meaningful part, and we're working to quantify it. Try the live HAL demo and rate a few AI answers at TrustChat.dev — the rating layer is being wired now, and your calls help train it. Building something similar? Tell us — let's build safe, ethical AI together.
Why now
Agents are already transacting; identity frameworks for AI are already being drafted — in legislatures, standards bodies, and the boardrooms of a handful of labs. Whatever gets built first becomes the default, and defaults are hard to unseat — and whoever lays the rails decides who is accountable, and whose provenance is trusted. The window to build the alternative — private, owned by the people, its chain of custody proven with zero-knowledge — is open now and won't stay open. That's why this is an open protocol, not a closed product. This is a direction we're building toward — a vision, honestly stated, not a finished product.
Come read the code, weigh in on the formulas, or just watch the build.
New here? Learn more — the fuller picture, plus a plain-language glossary →
Just the occasional update as the protocol takes shape — no spam, no selling your address.