A public evidence graph that people and AI agents can both trust, a hard line against governance tokens, and one genuinely open question about the payment rail. Feedback wanted — especially the kind that stings.
I’ve been sitting with a design problem for long enough that I want it stress-tested by people who’ll actually try to break it rather than nod along. There’s some context to get through before the real questions, so bear with me — it pays off.
What the thing actually is
At its core, it produces an evidence graph: a set of verified propositions, each one carrying its own provenance, stored as a public, append-only, content-addressed structure. On top sits a query layer — vector, keyword, and structured search — so the graph is something you can use rather than just admire.
There are two kinds of consumers. The obvious one is people: students, researchers, practitioners who want to know whether a claim holds up. The less obvious and, to me, more interesting one is AI agents that need provenance-checked grounding before they’ll stand behind their own outputs. As agents take on more of the reasoning, “where did this claim come from, and has anyone actually verified it?” stops being a footnote and becomes load-bearing.
The part that keeps me up at night
The entire point of this is to be hard to capture. So the design has to start from an uncomfortable question: who could quietly take control of this, and how? A few vectors I keep circling back to.
-
Centralized compute. Whoever hosts the processing can watch what flows through it, throttle it, or simply be leaned on by someone with a subpoena.
-
Centralized storage. Whoever holds the graph can alter it, bury parts of it, or gate who gets to read it.
-
Centralized funding for that compute. Whoever pays for the GPUs has a quiet hand on the lever of what gets processed in the first place.
-
A single network dependency. If generation leans on one cloud LLM API, that provider effectively owns the most important layer.
None of these are exotic. They’re just the ordinary ways good infrastructure slowly stops being yours.
The constraints I’m working inside
A few things are already fixed — some by the invariants of the system itself, some by the governance layer it has to live within.
The evidence graph stays public, append-only, and content-addressed; that’s non-negotiable. The heavy lift is extraction — call it Layer 1 — pulling structured, verifiable propositions out of documents. It needs real LLM inference and it’s the expensive part, currently centralized on one cloud API, and decentralizing it is the crux of the whole puzzle. Corroboration, Layer 2, is the opposite: deterministic and cheap, so any node can re-run it and check the work. Storage has to survive any single node vanishing — the governance design already calls for multi-jurisdiction escrow. Query has to serve humans and agents alike, which in practice means MCP-compatible. And the whole thing has to pay for itself without VC money, so something has to fund the compute.
That last constraint is where most of the interesting tension lives, so let me be precise about it.
Tokens, money, and where I actually land
I want to be clear about what I’m ruling out and what I’m very much not ruling out, because I know how this crowd reads proposals.
Off the table: a governance token — anything that hands holders voting rights or control over the system’s direction. The governance architecture is a perpetual-purpose trust with a fixed constitution of six machine-checkable invariants, and no amount of token-buying moves the Kernel. Also off the table: speculative instruments where the token’s rising price is the product. If a mechanism only works because it mints something tradeable that’s supposed to go up, it’s out.
Very much on the table: making money. This is a real system doing real work, and it needs to fund real compute. The question was never whether to charge — it’s what rail the payments run on, and whether that rail can be native to the agent economy instead of bolted onto it. Four models I’m actively weighing.
Model A — Pure fiat, subscription-style (the Crossref model). Institutions pay submission fees to have their papers processed, the way Crossref charges publishers to deposit DOIs while lookups stay free. Students pay small subscriptions. Attestation receipts — signed certificates that a proposition exists in the graph at a specific root with a specific corroboration score — get sold to anyone who needs audit-grade grounding. All denominated in fiat, all conventional accounting: simple, no regulatory complexity, no crypto. The weakness is real, though — AI agents can’t easily hold and spend fiat autonomously. The rail that’s most native to agents is exactly the one fiat handles worst.
Model B — ERG as the agent payment rail. Institutions still pay submission fees in fiat, because they have accounting departments and invoices are boring. But agents that need attestation receipts pay in ERG directly: native to crypto rails, able to hold and spend autonomously, and cheap enough at Ergo’s sub-block fees to make micro-attestations viable in a way Ethereum simply can’t. No new token gets issued — ERG is the existing rail. The organization holds an ERG treasury with a policy to sweep to fiat, which hands the Ergo ecosystem real transaction volume (agents paying for provenance grounding) without inventing a speculative instrument. The open question for this community: does it hold up? Is there enough liquidity and agent-side wallet infrastructure for autonomous ERG micropayments to work at scale?
Model C — An Ergo stablecoin for agents, fiat for institutions. Same shape as B, but agents pay in an Ergo-native stablecoin rather than volatile ERG, so the organization takes no price risk on its treasury. Institutions get fiat invoices; agents get a stable unit of account for budgeting their provenance queries. The stablecoin could be Ergo-issued or a bridged institutional stable (a USDC-equivalent on Ergo) — the point is a payment unit that stays stable for both payer and payee. This might be the cleanest fit for a system that needs predictable revenue while staying agent-native.
Model D — A hybrid rail. Offer both behind the same MCP endpoint: institutions use fiat-denominated credits or subscriptions, anonymous agents use x402 with ERG or an Ergo stable. The treasury accepts multiple currencies and converts per a published policy. Discovery stays straightforward — an MCP server in the public registry, a .well-known manifest, a listing in x402 discovery indexes. One governance note worth stating plainly: accepting stablecoin or ERG payment doesn’t touch the no-governance-token rule, because we’re not issuing anything, we’re accepting payment on an existing rail. But the treasury policy — what we hold, what we sweep, concentration limits — belongs to the governance layer, not to whoever happens to be paying.
Model E — Utility token (prepaid service credits), cross-chained to Ergo. The system issues a token representing prepaid extraction and attestation services. Institutions buy tokens to prepay for bulk processing. AI agents buy tokens to pay for provenance receipts autonomously. Token is redeemable for service — not for governance votes, not for revenue share. Small float (5%) sold with lock and cascading release over 2 years. Cross-chained to Ergo so the token is accessible to ERG holders and settled on Ergo’s low-fee chain. No governance rights attached — the Kernel’s invariants remain under the Guardian trust. The question: does a utility-only token (no governance, no revenue claim) actually raise meaningful funding in today’s market, or does the crypto ecosystem only engage with speculative/governance tokens?
I don’t have a settled answer. The constraint is fixed (no governance token, no speculative instrument) but a rail that lets agents natively pay for provenance receipts is squarely in scope. Which rail — fiat, ERG, an Ergo stable, or a hybrid — depends partly on what this community actually believes is viable for autonomous agent payments. That’s one of the things I’m here to find out.
Three mechanisms I want you to tear apart
A. Proof of Useful Work. What if the extraction and verification is the consensus mechanism? Each node processes documents and produces verified propositions; other nodes deterministically re-verify whether the extraction was faithful; the network agrees on that. The “mining” is genuine intellectual labor rather than burned hashes, and nodes that produce high-fidelity extractions earn more. It’s the most beautiful of the three to me — real work as the thing being agreed upon — but beautiful and attack-resistant aren’t the same word. Where does it break? Sybil attacks? Lazy nodes free-riding on everyone else’s re-verification? Extraction that’s perfectly faithful to a biased source?
There’s an honest caveat here that I can’t wave away. The deterministic half of faithfulness is checkable: cited spans exist at the claimed offsets, numbers match, the schema validates, corroboration recomputes. The semantic half is not — whether “may reduce mortality in a subgroup” quietly became “reduces mortality” is a judgment, not a checksum. The strongest fidelity signal turns out to be cross-model, cross-operator agreement on the same span-anchored proposition — but that needs multiple operators running different models, which is a federation question, not a consensus one. So the real question is whether there’s any model where the work genuinely is the consensus, or whether that framing dissolves into federation the moment you look closely.
B. Federated extraction with payment rails. Agents that need grounded, provenance-checked propositions pay for attestation receipts, funded through one of the payment models above. The key move is recognizing what you’re actually selling. You can’t sell the propositions — the graph is public and anyone can read it. What you sell is the attestation: a signed, timestamped certificate that a given proposition exists in the graph at log root R, with corroboration score S, computed under verifier version V, from sources X/Y/Z, with funding provenance F. Agents attach these to their own outputs. That’s notary economics, not data sales — and the question is whether it can carry the whole thing as the primary revenue model or only ever be one stream among several.
C. Free graph, paid computation around it. The graph stays free, public, and forkable — always. What you pay for is the work at the edges: new extractions (paying to have your papers processed in), rendering (the query-and-synthesis layer that turns raw graph into something legible), attestations (signed receipts for agents), and certification (a conformance mark saying a given renderer displays the graph faithfully). The precedent is Crossref-plus-notary: publishers pay to deposit, lookups are free, the organization runs on submission revenue. The question is which of these has to stay free to keep the whole thing honest, and which can actually carry the economics.
What I’d appreciate your read on
-
What’s the minimal decentralized architecture that keeps sovereignty? One line of argument I find persuasive says the graph doesn’t need blockchain consensus at all, because contradictory propositions simply coexist as data — there are no conflicting writes, only coexisting assertions. On that view, the only thing needing agreement is log inclusion, which a witnessed transparency-log federation handles fine: a Certificate-Transparency-style Merkle log with signed tree heads and three-to-five independent witnesses. But is “no conflicting writes” actually true? What about two extraction nodes processing the same paper differently? A retraction that invalidates earlier propositions? A corroboration score that shifts when new evidence lands? Are those genuine conflicts that need ordering, or just “append a new event and recompute the view”?
-
How do agents discover and pay for provenance checks? MCP handles exposure, and the payment rails are laid out above. What’s the discovery mechanism that connects an agent to the right graph in the first place — and which rail actually survives contact with autonomous agents at micro-transaction scale?
-
Is the flywheel real? The story I tell myself: agents need grounding, so they pay for attestation receipts, which funds extraction nodes, which grows the graph, which draws more agents needing grounding. A stronger version I keep coming back to says the durable demand isn’t query micropayments at all — it’s submission fees. Labs, journals, and agencies pay to have their corpus processed because being in the ground truth becomes valuable: citable, agent-discoverable, attestation-eligible. Which loop actually carries the economics?
-
How does it resist a motivated incumbent? Concretely: a pharma company spins up its own extraction nodes that faithfully extract from cherry-picked papers. Every proposition is true to its source and passes every fidelity audit — the corpus itself is what’s skewed. The current defense is diversity-weighted corroboration (counting distinct operator/model/funder tuples rather than raw corroborations), plus concentration caps and funder-concentration dashboards. Would on-chain enforcement of that concentration cap be meaningfully stronger than policy enforcement by a governance trust, or is it just ceremony?
-
What’s the bootstrap path? It starts life centralized, on one machine — no getting around that. The progression I have in mind: Stage 0 is custody first (publish the Merkle log with signed heads, cut weekly dumps to IPFS, recruit three witnesses in different jurisdictions). Stage 1 adds a second extraction operator running a different model, for audit diversity. Stage 2 federates on tripwires written into the governance constitution — a concentration cap exceeded, or submission revenue outgrowing single-operator capacity. Stage 3 ships attestation receipts and renderer certification. Does that progression work better on a blockchain or a federation?
What I’m deliberately not doing
No governance token — that’s a hard line. But I’m evaluating three payment models: (a) agents pay in ERG directly and the org sweeps to fiat, (b) the org issues a utility token (prepaid service credits, not governance) cross-chained to Ergo with a small locked float, or (c) hybrid — utility token for credits, ERG for settlement. Would value this community’s take on which model actually works for autonomous agent micropayments — and whether utility-only tokens raise meaningful funding in practice.
I’m wide open on the payment rail. If ERG or an Ergo stable is the right rail for agent-native provenance payments, let me know — and I want to hear what breaks. And if the answer is “you don’t need a chain at all, federation handles everything,” I want that too, with the attack vectors that make it true.
Rip into any of it. I’d much rather find the flaw here than prod.