# Claus beside IMD: what I would change in Design v0.1 (protocol and interface lens) Researcher 5 of 5, working alone. Desk research on public pages only, read on 2026-11-12. I did not log in, sign anything, or call any write or private endpoint. Every page cited is a public, unauthenticated GET. **Marking used throughout** - **[S]** read from a cited source (source key in brackets, list at the end) - **[I]** my inference from those sources - **[P]** my own proposal - **[U]** unverified: I could not confirm it from a public source Both networks change week to week. Claus's own docs say live capability endpoints take precedence over prose [C-llms]. Treat every number below as true on the read date only. --- ## 0. Corrections first: where v0.1 rests on a wrong or outdated fact | v0.1 says | What the sources say | Effect on the design | |---|---|---| | Claus has "a 2% pool fee, half to buy and burn CLAUS, half to compute and hosting" | **[S]** On the official Robinhood pool, the 2% project fee goes *entirely to liquidity*: "No current treasury share". On Ethereum, a 2.3% project fee accrues as ETH to one fixed ProjectRevenue recipient (CLAUS DEV). "New NFT, LP reward and buyback allocations are paused." The current priority is "build Robinhood liquidity with the complete 2% project fee" [C-llms]. The FOMO buyback hook is "retired" [C-llms]. | **[I]** Today no fee stream exists to pay seats, burn or compute. Item (4)'s "stated split between burn, seats and treasury" can only come from the request price, not from pool fees. | | Claus sells "x402 services in USDG" | **[S]** Claus's own paid services are priced in "USDC micro-units on Base" (`commerce_products`), and Contract Control "accepts direct x402 payments" in USDC on Base [C-openapi, C-services]. The service registry compares "ETH prices" (`public_service_compare`) [C-openapi]. Sponsored NFT inference "settles in USDC" over an x402 route on Base [C-nftinf, C-nftcap]. USDG shows up only in *third-party* City projects, for example Undertow: "1 USDG over x402 v2" [C-prodref]. | **[I]** Claus has no single settlement asset today. It uses USDC on Base for commerce, ETH or CLAUS for Jobs escrow, and CLAUS on Robinhood for the token. "One posted price in CLAUS over x402" would be new infrastructure, not a re-labelling. | | "395 API operations plus MCP" | **[S]** The OpenAPI document I fetched (version 0.41.0) has 426 operations [C-openapi]. Release status, dated 26 Oct, says "351 documented HTTP operations" [C-release]. | Minor. **[P]** Don't put a count in the design. Point to the schema. | | "Gap: no independent acceptance or onchain record" | **[S]** Claus already has: a `reviewPolicy` (reviewer and criteria) committed before a claim; `reviewer_candidates`, which "flag known shared ownership"; `app_task_review_submit`, an immutable per-criterion report whose author "cannot be a known controller of either party"; public objections; reviewer commissions paid "regardless of its conclusion"; pinned ETH or CLAUS escrow "deals" with a "single subsidized review case" and a signed "Claus decision" (`deal_review_request`, `public_deal_case`); and hash-addressed public deal evidence [C-openapi, C-econ]. Claus also writes `claus-journal/1` release records onchain on Ethereum [C-llms]. | **[I]** The gap is narrower than v0.1 says. Claus has independent review and arbitration *primitives*. What it lacks: (a) acceptance stays requester-only (`app_task_accept`), (b) review results are not written to any **portable** onchain reputation, (c) no planner routes one job across many seats. Build on what exists rather than design a new acceptance system. | | "No coordinator" | **[S]** A `coordinator` scope preset exists in `team_roles`, but it "contain[s] no financial or contract-administration permission" [C-econ]. | **[I]** True in substance. A coordinator *role* exists. A job-splitting control plane like IMD's does not. | | "No posted price" | **[S]** Work profiles carry an optional quote in ETH or CLAUS ("offers, not funded agreements"). Services return quotes. Commerce products have fixed USDC prices [C-econ, C-openapi]. | **[I]** There are prices, but per agent and per service. No network-wide price per request. | | The 300 houses can become seats as they are | **[S]** Each Claus NFT "represents 50,000 deposited $CLAUS; the current owner can burn it to redeem those tokens". NFTs, backing and stakes "remain on Ethereum" [C-llms]. Sponsored inference is "up to $5 per NFT per day" on `claude-haiku-5.5`, with 1,024 max output tokens, 6 requests a minute, concurrency 1, and `supports300Independent: false` [C-nftinf, C-nftcap]. | **[I]** (1) A seat can be destroyed by redemption, so seat identity must survive or cleanly end on burn. (2) Sponsored inference cannot do upgrade review: the model is small, output is capped, and running 300 at once is unconfirmed. v0.1 is right to require "the holder's own plan". Keep that, and say it is a hard requirement. | | ERC-8004 is a shared base on Robinhood Chain | **[S]** IMD lists ERC-8004 Identity and Reputation registries ("Not deployed by IMD") on Ethereum mainnet and Sepolia only. Its Robinhood Chain table lists none [I-docs]. ERC-8004 itself is still a **Draft** [E-8004]. Claus's public pages and OpenAPI make no mention of ERC-8004 [C-openapi, C-llms] (searched, none found). | **[I]** The shared identity layer is Ethereum mainnet. The Claus NFTs live there too, which is convenient. Payments happen on Robinhood Chain (IMD intake) and Base (Claus commerce), so records and payments are on different chains. | | CLAUS token governance (not in v0.1, but it matters for job #1) | **[S]** "The Robinhood token, hook and liquidity management are owner-upgradeable. The original developer is a single admin wallet with extra mint authority and no onchain mint cap or timelock. The public commitment is at least 1 hour of notice … It is not enforced by the contracts." [C-llms] | **[I]** This strongly supports "upgrade review" as the first job. It also means a review is advisory: nothing onchain makes the admin wait for it (see change 5). | | Thesis source: "Matt Huang's post of Nov 12 2026" | **[U]** A web search found no public text using "swarm riders" or "swarm native companies" [W-search]. | **[P]** Link the post in the design, or present the thesis as the proposer's own. The design does not depend on who said it. | Facts in v0.1 that I could confirm **[S]**: IMD charges a flat 0.5 IMD per request with no refunds [I-skill, I-docs]. IMD takes payment over x402 with Permit2, or through an intake contract with callback on Ethereum and Robinhood Chain (same address `0xa43e…de82` on both) [I-skill, I-docs]. 2,000 IMD NFTs [I-docs]. 8% of each launched token goes to seats connected at admission and 2% to wallets with accepted work, through a Merkle distributor [I-docs]. Claus Jobs claims last one hour, and only the requester accepts [C-jobs, C-openapi]. Agent mandates "expire within 30 days" [C-build]. New project inference is disabled [C-services]. IMD verdicts are written to the ERC-8004 Reputation Registry through IMD's WorkRegistry, "one per node per machine" [I-docs]. --- ## 1. Answers to the LENS questions ### 1.1 The smallest interface two swarms need to hire each other **[P]** Five messages and one read. Everything else is a convenience. 1. **Discover** (static): the swarm card (§1.2). It says what is sold, the price, the asset, the endpoint and which keys sign results. 2. **Quote**: the buyer sends a brief and gets back a price, an expiry and a `quoteHash` that commits to the exact body. Both sides already have this. IMD: `POST /requests/check`, then `/requests/quote`, free, 10-minute expiry [I-skill]. Claus: `service_quote`, then `service_accept` with "exact unexpired quote" [C-openapi]. 3. **Pay**: the buyer pays against the `quoteHash`. IMD: x402 v2 with Permit2, or `Intake.request` / `pay(quoteHash,…)` onchain [I-docs]. Claus: direct x402 USDC on Base for commerce, or a funded escrow deal for Jobs [C-services, C-openapi]. 4. **Receipt**: the seller returns `{requestId, resultHash, uri, signer, signature}`. IMD already returns `(requestId, resultHash, uri)`, but **[S]** "A job result: Not signed by the oracle … that they describe your work is the writer's word" [I-docs]. **[P]** Both swarms should sign receipts with a key named in the card (EIP-712), so a third swarm can check them without trusting either API. 5. **Review**: an independent party posts a verdict on `resultHash`. **[P]** Write it as an ERC-8004 `giveFeedback` (with `feedbackHash` = hash of the review document) or a `validationResponse` (see §1.3). 6. **Read** (status): one public GET per order. IMD has `GET /intake/:chain/:tx` and `/jobs/:id` [I-skill]. Claus has `service_order` and `public_review` [C-openapi]. **[I]** Both swarms already implement messages 2, 3 and 6, in incompatible shapes. The real interop work is a thin **adapter** per side plus agreement on 4 and 5. A new bridge is not needed. Settlement stays native to the seller (seller's chain, seller's asset), and the buyer adapts. ### 1.2 Is the six-question swarm card enough? **[I]** The six questions are the right human headings. The card is still missing what a *machine* or a *third swarm* needs to act without a phone call. **[P]** Keep the six, make the card JSON at a well-known URL, and add five fields: | Add | Why | |---|---| | **Identity anchor**: registry address, chain and agent-id range, or the adapter contract that binds seats | Lets a buyer check "who are your seats" onchain. **[S]** ERC-8004 already defines `registrations` and domain proof via `/.well-known/agent-registration.json` [E-8004]. | | **Endpoints and schemas**: OpenAPI or MCP URL, action ids and versions | **[S]** IMD versions actions (`job.open@job-1`) [I-skill]. Claus versions services through immutable manifests [C-openapi]. | | **Keys and trust assumptions**: which key signs receipts, which signs reviews, what each can and cannot do | **[S]** IMD already writes this in plain words: the writer key "cannot forge an answer but can withhold one" [I-docs]. Copy that style. | | **When it goes wrong**: refunds, disputes, liveness | **[S]** IMD: no refunds. Claus commerce: refund possible within seven days. Claus deals: one subsidized review case [I-docs, C-services, C-openapi]. These differ, so the buyer must see them before paying. | | **Independence policy**: how reviewers are chosen and which conflicts are excluded | **[S]** Claus flags shared ownership and bars controllers from reviewing [C-openapi]. IMD's verifier judges "independently of the agent that produced it" [I-docs]. | Also needed: `version`, `updatedAt` and `capacity`. **[S]** Claus already publishes capacity honestly ("simultaneous 300-request execution is not confirmed") [C-prodref]. **[I]** Prior art: the A2A protocol publishes an Agent Card at `/.well-known/agent-card.json` [A2A]. ERC-8004 registration files carry `services` and `supportedTrust` [E-8004]. **[P]** Make the swarm card an ERC-8004-style registration file *for the swarm*, served at `/.well-known/swarm-card.json` on imd.fun and claus.si, with a `registrations` entry linking it onchain. Don't invent a third format. ### 1.3 Is ERC-8004 the right shared identity, and what exactly is registered? **[I]** Yes for seats. No for riders as *agents*. Treat it as a Draft dependency [E-8004]. What I would register **[P]**: - **Seat (Claus house)**: one ERC-8004 agent id on the Ethereum mainnet IdentityRegistry, bound to the house `tokenId` by a Claus adapter, the same pattern as IMD's "ERC-8004 adapter … Binds an IMD NFT to its agent id" [I-docs]. Set `agentURI` to a registration file listing the seat's endpoints and `supportedTrust: ["reputation","validation"]`. Set `agentWallet` to the operator's hot wallet. **[S]** `agentWallet` is cleared automatically on transfer and needs a fresh signature [E-8004]. That fits Claus's rule that a transfer stops house edits [C-nftinf]. On **burn or redemption**, the adapter marks the agent inactive in metadata. The history stays [P]. - **Reviews**: ERC-8004 Reputation `giveFeedback(agentId, value, …, tag1="upgrade-review", tag2=, feedbackURI, feedbackHash)`. **[S]** `getSummary` *requires* a `clientAddresses` filter, because unfiltered results "are subject to Sybil/spam attacks" [E-8004]. IMD writes through its WorkRegistry, so IMD's record is `getSummary(agentId, [WorkRegistry])` [I-docs]. **[P]** Claus writes through one Claus review contract. A buyer who trusts both reads `getSummary(agentId, [IMD WorkRegistry, Claus ReviewRegistry])`. That one line is the whole cross-swarm reputation design. - **Independent second run**: use the ERC-8004 **Validation Registry** (`validationRequest` / `validationResponse`, with progressive states via `tag`) [E-8004]. The validator address is the *other* swarm's review contract. **[I]** IMD lists no Validation Registry deployment [I-docs], so this needs a deployment or an agreed address. That is an open question (§5). - **Rider**: **[P]** do not register riders as agents. A rider is the *client* in ERC-8004 terms: the address that orders and accepts. A rider's record is derived from receipts they paid for and accepted (IMD's `paidBy`, Claus's buyer). Each swarm can publish it as a signed, append-only document per rider address. Only add an onchain "rider registry" if priority or fee share ever has to be enforced by contract. ### 1.4 How independent acceptance should work on Claus **[P]** Don't build new machinery. Wire up what exists: 1. The requester commits `reviewPolicy` (reviewer id and criteria) *before* the claim, picking the reviewer from `reviewer_candidates` with ownership flags shown [C-openapi, C-econ]. 2. The seat delivers. The requester freezes the delivery (`app_task_review_request`). The reviewer submits one immutable per-criterion report. The reviewer is paid through a separate commission whatever the verdict [C-openapi, C-econ]. 3. **New** **[P]**: for upgrade review, also require a **second independent run**, meaning a second seat with no shared ownership flag reviewing the same frozen hash, and treat disagreement as `insufficient_evidence` [C-openapi]. Above a value threshold, the second run comes from **IMD** (an IMD `audit` job on the same repo and commit [I-skill]). 4. Acceptance stays with the requester (`app_task_accept`). It becomes *recorded acceptance* only when the reports exist. Disputes go to the existing deal review case, where "a returned resolution is an expiring signature" the parties submit themselves [C-openapi]. 5. **New** **[P]**: a small Claus relayer writes each finished report hash to ERC-8004 (§1.3). That fills the one real gap: no portable onchain record. **[I]** Reports "are advisory; they cannot sign for the buyer or pause escrow deadlines" [C-econ]. That is the right design and should stay. Independence is about the *record*, not taking control from the buyer. ### 1.5 Cleanest IMD → Claus and Claus → IMD calls, and what each side must trust **IMD workflow → paid Claus service** **[P]**: use **x402 directly** on a Claus service endpoint (Claus Contract Control already takes direct x402 USDC on Base [C-services]). Don't give IMD a Claus project credential: mandates are scoped and expire within 30 days [C-build], and service ordering via `service_quote` needs `service.call` on a Claus project [C-openapi]. That would mean IMD's control plane holding a long-lived Claus secret. With x402, IMD holds only a Base wallet with a spending cap. - IMD must trust: Claus's receipt-signing key (to be published in the card), and that Claus delivers. **[S]** Claus says "only a matching, confirmed USDC transfer unlocks" a report prepared before payment, with refunds possible within seven days [C-services]. - Claus must trust nothing about IMD keys. The x402 payment settles before delivery. **Claus app or hook → IMD job** **[S]**: a contract calls `Intake.requestWithFailure(action, body, callback, onFailure, IMD, 0.5e18)` on Robinhood Chain. The result comes back through the callback, `(requestId, resultHash, uri)`, with 200,000 gas, once, inside a try [I-docs]. - Claus must trust IMD's **writer key**. IMD states it can withhold a result, mispair it with a request, or send a false failure. It **cannot** forge an oracle answer. Job results are unsigned [I-docs]. **[P]** So a Claus hook must never move value on a job callback. If a hook needs a yes or no (for example "did the upgrade review pass?"), ask the **IMD oracle** a bound question and check `questionHash`. IMD documents exactly this pattern with "Did the review of request … pass?" [I-docs]. - Claus must hold **IMD tokens on Robinhood Chain** to pay. **[I]** That means the Claus treasury holds a second token. The founders need to decide this; it does not break the protocol. - IMD must trust nothing about Claus keys. Payment is pulled in the same call [I-docs]. ### 1.6 How others solved interop: what to borrow, what to avoid - **Borrow** A2A's well-known card and task lifecycle as the discovery shape [A2A]. Borrow ERC-8004's three registries and its insistence on client-filtered reputation [E-8004]. Borrow x402's "trust minimizing" rule, that the facilitator cannot move funds beyond client intent, and its explicit facilitator choice for mainnet [X402]. Borrow MCP, which Claus already exposes, as the *agent-facing* call surface [C-llms]. Borrow IMD's plain trust statements and its `quoteHash` binding [I-docs]. - **Avoid** a shared token or token bridge. **[S]** Claus deliberately migrated with "No CCIP" and "not … a shared CCIP bridge" [C-llms], and v0.1 already says no shared token. Keep it that way. **Avoid** unfiltered reputation scores (Sybil) [E-8004]. **Avoid** acting on unsigned results [I-docs]. **Avoid** promising that a review gates an upgrade the contracts don't gate [C-llms]. --- ## 2. Five changes to v0.1, ranked ### Change 1: Replace "bridge both ways" with a five-message protocol on existing endpoints, and make the card machine-readable - **What** **[P]**: define Quote, Pay, Receipt, Review and Read (§1.1). Each side writes one adapter over its existing quote and pay flow. The card becomes `/.well-known/swarm-card.json` with the six questions plus identity anchor, endpoints, keys, failure terms and independence policy (§1.2). Receipts are EIP-712 signed by a key named in the card. - **Why**: v0.1 item (5) describes three special-case integrations (intake orders, hosting calls, value-gated cross review). None of them generalises to a third swarm, which is the design's own stated requirement. **[I]** Both swarms already have quote, pay and status. Only signed receipts and a shared review format are missing. - **Cost**: one spec (a few pages), two adapters, and a signing key per swarm that must be guarded and rotated. IMD would need to sign job receipts it currently leaves unsigned [I-docs]. ### Change 2: Write reviews to ERC-8004 per swarm, and read them together with a client filter - **What** **[P]**: Claus houses get ERC-8004 ids on Ethereum mainnet through an adapter (§1.3). Each swarm writes verdicts from *one* known contract address. Cross-swarm reputation is `getSummary(agentId, [IMD WorkRegistry, Claus ReviewRegistry])`. The second independent run uses the Validation Registry, with the other swarm as validator. - **Why**: v0.1 says "written to the seat's record" without naming the record. **[S]** IMD already writes to ERC-8004 reputation through WorkRegistry [I-docs]. Claus has review reports but nothing portable [C-openapi]. ERC-8004 makes the client filter mandatory [E-8004], and the filter is exactly the "which swarm checked it" question on the card. - **Cost**: an adapter contract plus audit (a natural first IMD job ordered by Claus), Ethereum gas per batch (IMD batches up to 64 reviews per receipt [I-docs]; copy that), a Validation Registry deployment nobody has listed yet, and a dependency on a Draft standard. ### Change 3: Fix the money: native settlement per swarm, seats paid from the price, no burn split until there is revenue - **What** **[P]**: drop "half burn, half compute" (it is wrong, see §0). Each swarm sells in what it already settles: IMD in IMD, Claus in USDC on Base today. Claus can add CLAUS on Robinhood as a second accepted asset in its card when its x402 endpoint supports it. The posted Claus price per upgrade review covers (a) the seat, (b) the independent reviewer, paid whatever the verdict, (c) a treasury share. Publish the split only once real requests exist. The rider pays the *first* swarm, which subcontracts and pays the second in that swarm's asset. - **Why**: the Robinhood fee is 100% liquidity and the Ethereum fee goes to one fixed recipient [C-llms]. Nothing remains to split. Pricing in CLAUS-over-x402 assumes infrastructure Claus doesn't run; its x402 is USDC on Base [C-services, C-openapi]. "One payment" for a rider only works if one swarm takes the currency risk of buying the other's work. - **Cost**: weaker "every job burns CLAUS" story. The ordering swarm carries inventory of the other's token (Claus holding IMD on Robinhood Chain). Some price risk on subcontracted work. ### Change 4: Make seats real before routing work: holder's own plan, bound identity, honest capacity - **What** **[P]**: a house becomes a seat only when its holder (a) registers the ERC-8004 id via the adapter, (b) runs a worker on their *own* model plan, (c) passes a paid trial review. Sponsored $5/day inference stays a house perk, not a seat runtime. Burn or redemption retires the seat. Transfer clears `agentWallet` [E-8004]. - **Why**: sponsored inference is Haiku 5.5, 1,024 output tokens, 6 rpm, concurrency 1, and 300 at once is unconfirmed [C-nftcap, C-prodref]. That is not enough for upgrade review. A house NFT can be burned to redeem 50,000 CLAUS [C-llms], so an identity tied only to the NFT can vanish under an open job. - **Cost**: fewer active seats at first. **[I]** The 30-seat pilot may struggle to fill from 300 houses if holders lack a paid plan. The burn and transfer cases need code in the adapter and the Claus app. ### Change 5: Scope the first job honestly: upgrade review inside an enforced notice window - **What** **[P]**: the first coordinated job is review of each change to the CLAUS Robinhood token, hook or liquidity manager. The review runs *inside* the announced notice period, and its report hash is cited in the `claus-journal/1` onchain release record [C-llms]. Phase 2 also includes one structural change: put the owner behind an onchain timelock of at least the announced notice. The contract can be built by IMD (`launch.open` with `onchain: evm_contracts` [I-skill]) and reviewed by Claus seats. - **Why**: today a single admin can upgrade or mint with no timelock, and the 1-hour notice "is not enforced by the contracts" [C-llms]. A review the admin can skip is theatre. One hour is also too short for a two-run independent review, plus an IMD cross-review above threshold (IMD results arrive "minutes to days" later [I-docs]). - **Cost**: the Claus founder gives up instant upgrade and mint power, which slows emergency fixes. The timelock needs its own audit. --- ## 3. What I would cut - **[P]** "Half to buy and burn CLAUS, half to compute and hosting". It is factually wrong today [C-llms]. - **[P]** "IMD workflows end with a paid Claus service call **to host the build**". IMD already delivers sites to IPFS under ENS [I-skill]. Paid hosting duplicates that and adds a dependency to every workflow. Keep only the paid *second review* call. City listing can be a free, optional publish. - **[P]** The rider **fee share and priority credit** before Phase 4. They need revenue (§0) and a Sybil-resistant rider record. Keep the record itself. - **[P]** "A decision on a second seat class" and "the first swarm native company" as Phase 4 deliverables. They are goals, not protocol work. Ownership of a company with "no staff" raises legal questions this lens can't answer (§5). - **[P]** The API operation count, and any fixed numbers copied from prose, in the design text. Link the live sources instead [C-llms]. ## 4. What I would keep exactly as is - **[P]** **No shared token. Each swarm is paid in its own.** This matches Claus's no-shared-bridge choice [C-llms] and keeps a third swarm cheap to add. - **[P]** **Division of labour**: IMD builds, Claus houses, operates and reviews. It fits what each already does (IMD's launch factories and audits [I-docs]; Claus's hosting, apps, review and deal cases [C-openapi]). - **[P]** **Seats on the holder's own machine and AI plan.** It is the only version compatible with Claus's sponsored-inference limits [C-nftcap] and with IMD's model [I-skill]. - **[P]** **Upgrade review as the first coordinated job.** The governance facts make it the most valuable job available [C-llms] (scope change in change 5). - **[P]** **Phase 0 and Phase 1**, "hire each other with what exists". Every call is possible today with public endpoints: an IMD job paid on Robinhood Chain [I-skill], a Claus service paid in USDC on Base [C-services]. - **[P]** **The six questions** as the card's human headings (extended in §1.2). ## 5. Open questions only the founders can answer 1. **Claus**: will you put the Robinhood token, hook and liquidity owner behind an onchain timelock? If not, is an advisory review still worth paying for? 2. **Claus**: which asset does a Claus seat get paid in: CLAUS on Robinhood, USDC on Base, or both? Will Claus hold IMD inventory to order IMD jobs? 3. **Claus**: is sponsored house inference meant to continue for houses that become seats, or does it end when a holder's own plan takes over? 4. **Claus**: when a seated house is burned to redeem its 50,000 CLAUS, what happens to open jobs and the record? Retire, transfer or freeze? 5. **IMD**: will IMD sign job receipts (today unsigned), and publish the signing key in a swarm card? 6. **IMD**: will IMD count a Claus review contract as a trusted `clientAddress` when reading reputation, and treat a Claus verdict as an independent second review inside its own pipeline? 7. **Both**: who deploys, or which address do you agree on for, an ERC-8004 Validation Registry on Ethereum mainnet? (None is listed by IMD [I-docs].) 8. **Both**: what is the value threshold above which the other swarm must review, and who pays when its verdict disagrees? 9. **Both**: who legally owns and is liable for a "swarm native company" with no staff? 10. **Proposer**: where is the Matt Huang post? I could not find it [W-search]. --- ## 6. Revised one-page design (v0.2, proposal only, nothing agreed) **Goal.** Claus is a second swarm beside IMD. It is not a copy, and its link to IMD must work unchanged for a third swarm. **Corrected facts it rests on.** Claus's Robinhood pool fee goes 100% to liquidity. Its Ethereum fee goes to one fixed recipient. Burn and buyback allocations are paused. Claus commerce settles in USDC on Base. CLAUS on Robinhood is owner-upgradeable with no timelock. The 300 houses are Ethereum NFTs, each burnable for 50,000 CLAUS. ERC-8004 registries in use are on Ethereum mainnet. IMD charges 0.5 IMD per request, no refunds, with unsigned job results and signed oracle answers. **Protocol (the part a third swarm adopts).** 1. *Card*: `/.well-known/swarm-card.json`, an ERC-8004-style registration file for the swarm. It answers who the seats are (registry and adapter), what is sold (action ids and schemas), price and asset, how to pay (x402 scheme and network, or intake contract), what comes back (signed receipt format), who checked it (review contract address and independence policy), failure terms (refunds, disputes, liveness) and keys (signer addresses and powers). 2. *Five messages*: Quote (bound by `quoteHash`), Pay (seller's chain and asset), Receipt (EIP-712 signed `{requestId, resultHash, uri}`), Review (ERC-8004 feedback or validation on `resultHash`), Read (public status GET). 3. *Reputation*: each swarm writes from one contract. Readers filter by the swarms they trust: `getSummary(agentId, [swarmA, swarmB, …])`. 4. *Calls between the two*: IMD → Claus over x402 with no shared credential. Claus → IMD through the Robinhood intake with callback. Hooks act only on bound oracle answers, never on unsigned job callbacks. **Claus work (in order).** (a) Adapter binding house to ERC-8004 id, built by IMD and reviewed by Claus. (b) Seat admission: own plan, trial review. (c) Upgrade review on existing `reviewPolicy`, reviewer reports and deal cases, plus a second independent run, with report hashes in the onchain journal and in ERC-8004. (d) One posted Claus price per review covering seat, reviewer and treasury. **Phases.** 0: founders answer §5 Q1, Q2, Q5 and Q7; yes or no. 1: hire each other with what exists (three IMD jobs ordered by Claus, starting with the adapter and timelock; three Claus reviews ordered by IMD over x402). 2: 30-seat upgrade-review pilot behind the timelock; signed receipts both ways. 3: cards live on both domains; cross-swarm review above the agreed threshold; rider record published as signed per-address documents. 4: all seats that qualify; card opened to a third swarm; rider fee share and a swarm native company only if Phase 3 shows revenue. **Not in scope.** A shared token or bridge, paid hosting inside IMD workflows, and fee shares paid out of pool fees that don't exist. --- ## Sources (all fetched 2026-11-12, public GET only) - **[C-llms]** Claus, llms.txt: https://claus.si/llms.txt (migration and fees, governance, NFTs, onchain journal, Jobs, MCP) - **[C-openapi]** Claus OpenAPI 0.41.0: https://claus.si/__network/openapi.json (`app_task_review_*`, `reviewer_candidates`, `deal_*`, `service_*`, `commerce_*`, `team_roles`) - **[C-econ]** Claus, Work in Claus Network: https://claus.si/docs/agent-economy - **[C-jobs]** Claus Jobs guide: https://claus.si/docs/jobs - **[C-build]** Claus agent start guide: https://claus.si/build.txt - **[C-services]** Claus Services: https://claus.si/services - **[C-nftinf]** Claus NFT inference: https://claus.si/docs/nft-inference - **[C-nftcap]** Claus NFT capabilities (live JSON): https://claus.si/__network/v1/nft/capabilities - **[C-prodref]** Claus product reference: https://claus.si/product-reference.json - **[C-release]** Claus release status: https://claus.si/docs/release-status - **[I-skill]** IMD skill: https://imd.fun/skill.md - **[I-docs]** IMD docs (full): https://imd.fun/docs/llms.txt (intake, callbacks, "What a callback proves", contract tables, WorkRegistry, swarm share) - **[E-8004]** ERC-8004: Trustless Agents (Draft): https://eips.ethereum.org/EIPS/eip-8004 - **[X402]** x402 README: https://github.com/coinbase/x402 (raw: https://raw.githubusercontent.com/coinbase/x402/main/README.md) - **[A2A]** A2A protocol specification: https://a2a-protocol.org/latest/specification/ - **[W-search]** Web search "Matt Huang swarm riders swarm native companies", 2026-11-12: no matching source found (results included https://www.paradigm.xyz/team/matthuang and https://www.antoinebuteau.com/lessons-from-matt-huang/, neither using the terms) **Limits.** These are desk reads of public pages only. I could not see live job volumes, the contents of private reviews, or whether any endpoint performs as described. Prose and live JSON were sometimes dated differently (for example, the operation counts), and I recorded both. Nothing here was checked by an independent reviewer.