← hyperdag.org

The fuller picture

The longer read — and the words we use, in plain language.

The front page is the short version. This is the long one: how we keep AI honest, the trust you own, real privacy, and the architecture underneath. Honest definitions, not marketing — where something is designed but not yet shipped, we say so. Every term has its own anchor, so any page can link straight to it.

Building on the protocol rather than reading about it? TrustShell's glossary defines the terms you meet in the tools — RepID, grants, stake, and the MEASURED / APPROXIMATE / NOT_CHECKED / FAILED vocabulary — in a developer-facing frame, and links back here for the protocol depth. Two glossaries, each definitive for its own surface, neither maintaining a copy of the other's.

Group one

How we keep AI honest.

The checks that stop a confident-sounding answer from being taken on faith.

SBFA — Stochastic Bias Fracture Array

We never let AI models from the same family check each other's work — same family, same blind spots. So we deliberately spread the work across independent, decorrelated validators, where a bias in one isn't shared by the next. A flaw that would slip past a room of similar models gets caught by one that fails differently.

The established idea it names: avoiding "common-mode failure" — engineering's term for when redundant parts all break the same way at the same time, so redundancy buys you nothing.

HAL — Hallucination Assessment Layer

A checkpoint that flags claims an AI can't ground in evidence — before those claims count for anything. If a statement can't be traced back to a verifiable source, HAL marks it rather than letting it pass as fact.

Grounding

Tying every claim to a verifiable source — the opposite of "just trust me." A grounded answer can show its receipts; an ungrounded one is asking for faith.

Drift

When a model quietly changes behavior over time — answering differently this month than last, without anyone deciding it should. We measure drift so it can't creep in unnoticed; a change you can see is a change you can question.

Comma Veto Pythagorean comma

When validators agree too confidently, that can signal a shared blind spot rather than the truth — so unusually high-confidence agreement can trigger a second look instead of a rubber stamp. Suspiciously perfect consensus is treated as a question, not an answer.

Named for the Pythagorean comma (531441 ⁄ 524288 ≈ 1.0136) — the tiny musical dissonance that reveals a scale is subtly out of tune. The small mismatch is the tell.

BFT — Byzantine Fault Tolerance

Reaching the right answer even when some validators are wrong, offline, or actively malicious. Named for the "Byzantine generals" problem in computer science — the classic question of how to agree on a decision when you can't fully trust every participant.

Blind adjudication

Validators judge without seeing one another's verdicts — so they can't quietly rubber-stamp the crowd. Each reaches its own conclusion first; agreement means something only because it wasn't copied.

Anti-fragile

Not merely robust — it gets stronger under attack. Red-teaming hardens it, bad actors get slashed, and stress reveals weaknesses so they can be fixed. Where a fragile system breaks under pressure and a robust one endures, an anti-fragile one improves because of it.

Paper trade / testnet test tokens

Sandbox activity with test tokens on a testnet — the safe place to learn, train, and test your agents before anything real is on the line. No real money is at risk. The behavior is real; the stakes are not, so a mistake here costs nothing but teaches everything.

Goodhart's Law

When a measure becomes a target, it stops being a good measure — people optimize the number, not the thing it was meant to track. We build against it: no single score rules, validators can't grade their own family, and gaming one signal doesn't move the others.

Named for economist Charles Goodhart; Campbell's Law, the Cobra Effect, and the Lucas Critique are the same trap in other fields.

CRAG — relevance-graded retrieval

Before an AI answers from retrieved sources, we grade whether those sources are actually relevant and sufficient — and re-retrieve or correct when they're not. A confident answer built on the wrong passages gets caught before it's given.

CRAG = Corrective Retrieval-Augmented Generation.

Decorrelated validator

A checker chosen precisely because it fails differently from the others — different model family, different training, different blind spots. Correlated validators all miss the same errors; decorrelated ones catch what the rest can't.

RRL — Response Reputation Layer

A layer that scores each AI response for how trustworthy it turned out to be, and feeds that back — so reliable sources get weighted up and unreliable ones down over time. Reputation attached to answers, not just to accounts.

Proof-carrying retrieval

When an AI pulls a fact from memory, it returns the fact together with a cryptographic proof that the fact really is in the committed record — so you can verify the source wasn't fabricated or altered, without having to trust the retriever.

The retrieval carries its own receipt.

Prompt injection

An attack where hidden instructions buried in a web page, document, or tool result try to hijack an AI into doing something its owner never asked — leaking data, sending money, ignoring its own rules. Treating everything an AI reads as untrusted input is the defense.

The AI-era cousin of SQL injection.

Group two

Trust you own.

Reputation and receipts that belong to you — carried, not rented.

RepID

Reputation earned through honest, verified work — portable and yours. It accrues from what you (or your agent) actually do, and you carry it between apps and contexts instead of rebuilding it inside each walled garden. Not a score handed down by a platform; a credential you earn and hold.

Purpose Gate anti-theater

Reputation only moves on real deliverables — never for looking busy. Activity that produces nothing of substance doesn't earn (or lose) standing. The gate exists to keep the score honest: it measures work done, not motion made.

Staking your guardrail

Your guardrail. Staking caps an agent's blast radius — the most it can transact on its own — while letting it earn the right to do more. It's never an open checkbook, and never access to your credit card or crypto accounts. Safer and more empowering at once: you set the ceiling, your agent earns the trust to raise it.

A stake is denominated in USDC (a dollar-pegged stablecoin) or ETH, held on Ethereum / Base — real value you lock as your guardrail, not a fee we take.

Authority ceiling

The most an agent can transact on its own — set by your stake and your tier. Raise your stake, raise the ceiling, within an anti-whale curve so it stays fair: more stake buys more room, but with diminishing returns rather than letting the biggest wallet simply run away with it.

Provenance

The tamper-evident chain of custody — who did what, when, and under whose authority — verifiable end to end. When an agent acts on your behalf, provenance is the record that lets anyone check the trail rather than take it on trust.

Glass box (vs black box)

We can't honestly show you why a model thinks what it thinks — no one can. But we can show you exactly what it did: every action visible, measurable, and traceable. That's a glass box against the industry's black box — trust delivered as evidence, not asserted as a claim.

ERC-8004

The emerging on-chain standard for agent reputation — a public registry where an agent's earned trust is recorded on the blockchain, so it's portable and verifiable by anyone, rather than locked inside one company's database.

An Ethereum standard (ERC = Ethereum Request for Comments), the same kind of open spec as ERC-20 for tokens.

x402

A payment protocol that lets one agent pay another automatically over the web — reviving the long-dormant HTTP "402 Payment Required" status so machines can settle micro-payments for services without a human at the checkout.

Code 402 was reserved in the original web spec for exactly this, decades before agents existed to use it.

Portable earned trust

Reputation you carry across every app and vertical instead of rebuilding it inside each one. Trust earned in one place travels to the next — because it belongs to you, not to the platform that happened to witness it.

Bilateral learning loop

Every check scores both sides at once — the one making a claim and the one verifying it. Make a bad claim and get caught, your standing falls; catch a bad claim correctly, yours rises. Because both are scored from a single interaction, gaming both at the same time is hard — which is the whole point.

Spectrum of autonomy

How much an agent may do on its own — from suggest-only, to act-with-your-approval, to fully autonomous within limits. Autonomy isn't all-or-nothing: you grant it in degrees, and an agent earns more room as it proves trustworthy.

Wallet

Your keychain for the on-chain world — it holds your keys, your identity commitment, and any stake or tokens, and it's how you authorize what your agent may do. Bring one you already have, or create one the safe way as part of signing up.

Faucet

A free tap that hands out test tokens on a testnet, so you can practice, train agents, and try everything with zero real money at risk before you ever touch the real thing.

Called a faucet because the test tokens "drip" out on request.

Group three

Privacy that's real.

Prove what's true without handing over what's private.

ZKP — Zero-Knowledge Proof

Prove something is true without revealing the data behind it. You can show you're over 18, hold a credential, or are authorized to act — while revealing nothing else: no date of birth, no document, no raw data changing hands, only the proof that the claim holds. It's the cryptography that lets self-sovereignty be real instead of a slogan — verified without being surveilled.

Selective disclosure

Share only what you choose — nothing more. Instead of handing over a whole document to prove one fact from it, you reveal exactly the fact that's needed and keep the rest to yourself.

Self-sovereign identity

You hold the keys — no company or government owns your identity. Your credentials live with you and are presented on your terms, rather than being stored, controlled, and rented back to you by a platform.

Nullifier

Proves you're unique or eligible without revealing who you are. It lets a system confirm "this is one distinct, allowed participant — and hasn't already acted" while keeping the person behind it anonymous. The mechanism behind one-person-one-vote or no-double-spend, minus the surveillance.

Homomorphic encryption

Computation on data that stays encrypted the whole time — a service can process your information and hand back a correct result without ever being able to see the information itself. Your data stays useful without being exposed.

"Homomorphic" means the math done on the encrypted data mirrors the math on the real data, so the answer comes out right.

Group four

The architecture.

The ledger and learning underneath — what makes it fast, verifiable, and shared.

HyperDAG

Hyper (fast) + DAG (Directed Acyclic Graph — a horizontal, parallel ledger) — a hybrid of DAG and blockchain. Like a blockchain, but horizontal and parallel instead of linear and vertical, so many things can settle at once, with a blockchain underneath for final, tamper-proof settlement. Fast where it should be fast; immutable where it must be.

DAG / DLT

A DAG (Directed Acyclic Graph) is a ledger where transactions run in parallel — a branching web that never loops back on itself — rather than a single linear chain. DLT (Distributed Ledger Technology) is the umbrella term: any shared record kept in sync across many independent computers with no central owner. Blockchain and DAG are both kinds of DLT.

Why it matters here: a classic blockchain is linear and vertical, which caps its speed; a DAG is horizontal and parallel, which is what gives autonomous agents the throughput they need.

A2A / federated learning

Agents learn from each other — A2A is agent-to-agent — and the network's intelligence compounds as they do. In federated learning, that shared learning happens without pooling everyone's raw data in one place, so the value created stays with the people who created it rather than flowing to whoever owns the servers.

ANFIS — the router's learning brain

An Adaptive Neuro-Fuzzy Inference System. It learns, from every past outcome, which validators and models catch which kinds of error — and routes each new job to the ones most likely to get it right per dollar. Fuzzy logic's human-readable rules, tuned by a neural network's learning.

Adaptive Neuro-Fuzzy Inference System.

LASSO — keeps the router focused

A math technique that automatically finds the few signals that actually matter and zeroes out the rest — so the router pays attention to the handful of features that predict a good outcome instead of drowning in noise.

Least Absolute Shrinkage and Selection Operator — the "L1" penalty that drives weak features to exactly zero.

MoE — Mixture of Experts

Instead of one giant model doing everything, many specialized ones each handle what they're best at, with a router sending each task to the right expert. The right specialist beats a jack-of-all-trades — cheaper and better at once.

Mixture of Experts.

Adaptive RAG routing

Deciding, per question, how hard to work: a quick lookup for an easy one, deep multi-source retrieval for a hard one. The system matches its retrieval effort to the difficulty instead of spending the same on every query.

RAG = Retrieval-Augmented Generation — letting a model look things up instead of guessing from memory.

Durable dispatcher

The tireless clockwork that keeps the swarm moving — it hands out work, retries what fails, backs off when a provider is rate-limited, and never loses a task even if a worker dies. Persistent by design, so the loop survives restarts and outages.

Distributed fabric

The whole ecosystem seen as one woven cloth rather than separate apps — agents, validators, ledgers, and storage spread across many machines with no single owner, all sharing one substrate of trust. The verticals are patterns on the same fabric.

Personal AI OS

The long-term shape of it: your own trusted operating system for AI — your agents, your keys, your reputation, your data, under your control — instead of renting slices of someone else's. The phone put a computer in your pocket; this puts a trustworthy AI layer under your name.

← Back to hyperdag.org  ·  TrustShell glossary — the developer-facing terms →