The thinking behind Veritas — eleven ideas and seven research dives, in plain English.
Behind the working paper there are eighteen separate documents that explore specific design questions. Two paragraphs on each, here. If you want depth on any one, the technical version is one click away.
~ 10-minute scan · Pick what interests you
AThe eleven design ideas
Why we admit checkers who think each other are evil and lying.
The whole point of a "plural verdicts" system is that different communities check things differently — even communities that fundamentally disagree. If we only let in nice agreeable institutions, we've built a single-frame system with marketing. We have to admit checkers from communities that hate each other. That's the design choice.
This is uncomfortable. It means state-aligned validators, religious-tradition validators, dissident-community validators all coexist. The reader's profile decides whose attestations they see. The protocol records who said what; it doesn't pretend disagreement doesn't exist.
Chain for the slow part, federation for the fast part.
Underneath: a public blockchain (Ethereum Layer 2) that keeps the permanent record. Above it: a network of fast aggregators that read the chain, apply your profile, and serve answers in milliseconds. The chain handles "what was said and who paid whom." The federation handles "what to show this user right now."
Most failed fact-check projects put everything on the blockchain. That's slow and expensive. Most didn't have a permanent layer at all — they couldn't handle communities that don't trust each other. Hybrid uses each tool for the job it's good at.
How a token can pay validators without becoming a speculation casino.
The token is what AI companies and websites pay with when they buy services. Validators earn it for their work. The treasury holds the actual stable value (USDT, fiat, T-bills). Validators can convert tokens to stable assets via the treasury. This is the Chainlink survivor pattern — service payment, no mandatory burn, no speculation pump.
Several previous projects (Civil, Po.et, Bitpress) failed because they built speculation tokens that drove the wrong incentives. We're trying to learn from that. The next paper version may drop the burn-to-cash mechanism entirely on regulatory advice.
A small file that says which checkers you trust on which topics.
CPML stands for Consensus Profile Markup Language. It lives on your device. It encodes things like "for science questions trust the scientific consensus; for legal questions trust EU jurisprudence; for history show me everything." When you read a page, your profile tells the aggregator how to compose the verdict for you.
The technical model borrows from a 2003 academic framework called Value-based Argumentation Frameworks. The current sketch isn't quite a faithful implementation — the critical reviewer caught us on that. v0.3 either commits to the formal model or honestly drops the citation.
Six income streams, with one carrying most of the weight.
AI laboratories pay for grounding. Websites pay for badges. Users pay subscriptions. Content publishers pay for priority verification. Contested-claim parties pay for investigations. Foundations donate.
The honest version: the AI-laboratory revenue is the load-bearing pillar. If it doesn't materialise, the whole compensation model has to shrink. The next paper version (v0.3) tiers the projections honestly and presents a fallback model for the case where AI labs don't sign up.
Pay-to-investigate for contested claims — and a way to fund investigative journalism.
If two parties care strongly about a contested claim, they can both pay for a formal investigation. The money goes into escrow, qualified investigators take the case, they document their reasoning, sign their verdict, money releases. Both sides get a documented investigation; the public sees it.
This is the most economically novel piece. It scales verification effort to where money lives. It gives investigative journalists a direct revenue stream. It's also the riskiest mechanism — well-funded actors can game it. v0.3 specifies structural defences (flat-fee tiers, parity requirements, public-interest fund for under-resourced claims).
A Buzzfeed-style quiz that gives you your consensus profile and seeds the whole user base.
Most users won't write a CPML by hand. So the very first product we'd ship is a 10-question quiz: "for each of these statements, which standards of evidence would you accept?" The result: your starter profile. The mechanic: shareable, fun, takes 5 minutes.
Names being considered: Frame, Compass, Plural. The quiz works as a standalone product even before validators populate the system — and it bootstraps the user base for when the protocol goes live.
Different laws in different countries — same protocol, different filters.
Germany forbids Holocaust denial; the US protects most of the same speech. India has its own rules. We can't apply one filter to the whole world. Each country gets a chapter — an aggregator that complies with local law. The chain is universal; what's shown depends on which chapter's view you're using.
Wikipedia uses this pattern. ICANN uses something similar. It's a tested model. Day-one chapters: EU, US, UK, Switzerland. Others as partner availability and jurisdictional need emerge.
What the protocol refuses to record, and who decides.
Five operations are protocol-refused: verifying child sexual abuse material, verifying non-consensual intimate imagery, verifying credible imminent threats, verifying mass-casualty-weapon synthesis, verifying or locating illegal markets. The list is narrow on purpose — operation-based, not topic-based, so it can't drift into political content moderation.
A 9-seat multi-stakeholder dispute panel handles edge cases and revisions to the list. Adding to the list requires 7-of-9 supermajority + 60-day public comment. The reviewer caught us: items 3 and 4 are arguably topical, not operational. v0.3 redraws the line honestly — most of these move down to aggregator-level filters, not protocol refusals.
Which blockchain to build on. Probably Base. Possibly own.
Phase II will build on an existing Ethereum Layer 2 — most likely Base, with Optimism as a backup. Cheap (less than a cent per record), fast (sub-second confirmation), regulatory clarity good (Ethereum-ecosystem chain).
If at Phase III we hit specific blockers — jurisdictional neutrality issues, hostile-frame requirements that don't fit Base, crypto-investor preferences for own chain — we evaluate migrating to a "sovereign rollup" using Rollkit on Celestia. Default position: stay on the L2 indefinitely.
The argument that moved us from "no blockchain" to "yes blockchain."
An earlier draft of the working paper (v0.1) explicitly said: don't use blockchain — use federation only. v0.2 reverses that. Why? Because once you genuinely need to admit mutually-hostile validators, no neutral foundation can credentialise enemies. A blockchain is the right tool when participants don't trust the same authority.
This idea documents the four propositions that shifted the architecture. Plus four falsification conditions: things that, if observed, would reverse the decision again and put us back on federation-only. Honest record of the reasoning, kept so future contributors don't repeat the debate from scratch.
BThe seven research investigations
Beyond the design ideas, seven longer research dives by independent specialists. Each is 30-90 KB of careful work that the working paper draws on.
What the academic literature says about modelling user preference profiles for trust composition.
An 8,800-word review covering semantic-web preference vocabularies, argumentation frameworks with agent preferences, epistemic communities, moral foundations theory, filter-bubble research, modal logic, truth-maintenance systems, preference aggregation impossibility theorems, and the postmodern-truth philosophical critique. Recommends modelling profiles as Value-based Argumentation Framework "audiences" rather than political-compass coordinates.
What's already out there that does some of what Veritas wants to do.
30+ tools surveyed: political-compass tests, news-consensus aggregators (AllSides, Ground News, Tangle), deliberation platforms (Pol.is, Kialo), annotation systems (Hypothes.is), personality-quiz traction patterns, matchmaking apps, epistemology-adjacent products. Top three to copy: 16Personalities mechanic, Tangle's steelman-then-synthesise, iSideWith's share-card-as-core-product.
How the money model holds up against real-world precedents.
Examines Chainlink LINK, Witnet WIT, Kleros PNK, UMA, Arweave, prediction markets — the tokens that work and the ones that don't. Recommends dropping burn-to-USDT for Phase II launch (regulatory weakness), using flat-fee investigation tiers, KYB above $1K threshold. Year-3 revenue projections in three tiers; large gap between optimistic and base case acknowledged honestly.
How to build the technical stack with existing open-source pieces.
Recommends Base as primary chain + Optimism mirror, Ethereum Attestation Service as the attestation primitive, AT Protocol's label spec as the wire format, libp2p for gossip, IPFS for evidence storage. Cost target: ~$0.001 per attestation, ~$40K/month operational floor at 1M attestations/day. Closest structural precedents: Lens Protocol + Momoka, Bluesky Ozone labellers, Sigstore Rekor.
What the EU, US, UK, Singapore, UAE, Switzerland regimes mean for the project.
MiCA articles, SEC Howey test applied to utility tokens, FINMA / MAS / VARA / ADGM comparative analysis, Wikimedia chapter governance precedents, Creative Commons jurisdictional porting history, ICANN regional-organisation models. Recommends Swiss Stiftung + Swiss Verein + EU operating GmbH (Ireland or Netherlands or Malta) as the legal structure. Day-one chapters: EU-Brussels, US, UK, Switzerland.
What to fork, what to read, what to participate in.
Top 8 repositories to fork or borrow: Ethereum Attestation Service, Sigstore Rekor + Trillian-Tessera, IETF SCITT reference, Bluesky Ozone, Community Notes scoring algorithm, Pol.is math service, 8values quiz, OpenFactCheck + Vectara hallucination eval. Three specific developer-community pitfalls documented (token-weighted oracle capture per Polymarket/UMA 2025, cursor-based firehose label drops, general Merkle-tree transparency-log scaling).
How to actually build a viral consumer product without becoming a manipulation machine.
Case studies on viral-quiz mechanics (16Personalities, Political Compass, BuzzFeed, OkCupid, Spotify Wrapped, Pew typology). Ethical analysis of dark patterns. Matching-without-echo-chambers techniques (Pol.is, Bumble BFF). Opposite-view surfacing patterns (Ground News, Tangle, Braver Angels). Revenue mix analysis (Wikimedia $185M, Mozilla single-source trap, Brave $100M, Signal $25.8M, Filecoin RetroPGF). Final naming recommendation: Frame, Compass, or Plural.
CHow these connect
The eleven design ideas are choices that have to be made. The seven research investigations supplied the evidence behind those choices.
If you want to engage with the project at the design level — push back, propose alternatives, find holes — these eighteen documents are where the substantive thinking lives. The working paper is the synthesis; this is the source material.
The critical review found that some of the syntheses don't quite fit what the source documents actually say. v0.3 reconciles those.
ELI5 home · The proposal (plain) · What critics said (plain) · What we're fixing (plain) · Questions · All technical versions