# Claus as a swarm beside IMD: what to change in Design v0.1 Synthesis of five independent views (protocol and interface lens) · 12 November 2026 **Verdict [Proposal].** All five researchers reach the same answer by different routes. Run a bounded hiring pilot between the two swarms, but do not adopt v0.1 as the protocol. What the swarms share should be a versioned work agreement with separate receipts for payment, delivery, review and acceptance. A common seat model, a common token or a "bridge" should not be part of it. Claus already has most of the review machinery v0.1 says it lacks. What it lacks is an acceptance and authority policy, receipts that are portable across swarms, and payment rails that match the price v0.1 posts. The work should go there first. ## How to read this report - **[Read]**: a cited public source says this. That is attribution, not proof that the system behaves this way. Unless marked *(spot-check)*, the source was read by the views named. - **[Inference]**: a conclusion drawn from cited evidence. - **[Proposal]**: a design recommendation. Nothing here is agreed by either team. - **[Open]**: a question that only the founders, or the page maintainers, can settle. - **[Uncertain]**: the sources conflict or could not be checked. - **V1 to V5** are the five input views (artifacts/view-1.md to view-5.md). "5/5 (V1–V5)" means all five views propose the point. - **Synthesis spot-check.** While writing this report I fetched four public pages myself with plain GETs on 12 Nov 2026 to judge two conflicts between the views: [llms.txt](https://claus.si/llms.txt), [migration-reference.json](https://claus.si/migration-reference.json), [public capabilities](https://claus.si/__network/v1/public/capabilities) and the [inference guide](https://claus.si/docs/inference). Results are marked *(spot-check)*. I did not log in, sign, purchase or call any write endpoint. Every other source citation comes from the views. I did not re-verify those sources, except that the views corroborate one another where several cite the same page. - **Limits.** This is desk research only. No view ran a paid flow, measured uptime, audited contracts or tested whether operators are really independent. Live pages change. Numbers are as observed on 12 Nov 2026. --- ## 0. Factual corrections to v0.1 (made before any design change) | # | v0.1 says | What the sources say | Views | Status | |---|---|---|---|---| | F1 | Claus has "no independent acceptance or onchain record" | **[Read]** Claus Jobs already supports the following. A `reviewPolicy` (reviewer plus rubric) is committed before the claim. Input, output and evidence are frozen through `app_task_review_request`. The reviewer submits an immutable per-criterion report (`app_task_review_submit`) with pass, fail or `insufficient_evidence`. There are objections (`app_task_review_challenge`) and a separate review commission that is paid whatever the verdict. Reviewer matching (`reviewer_candidates`) flags known shared ownership and warns that "distinct identities do not establish independent operators". The docs say "Review is advisory", the report "Does not accept the task or release funds", and only the requester accepts (`app_task_accept`). Paid Jobs bind commitments to onchain escrow. Claus also writes `claus-journal/1` release records onchain. Sources: [Jobs](https://claus.si/docs/jobs), [OpenAPI](https://claus.si/__network/openapi.json), [Work in Claus Network](https://claus.si/docs/agent-economy), [llms.txt](https://claus.si/llms.txt), pinned [reviews.mjs](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/src/apps/reviews.mjs). | 5/5 | **Corrected.** **[Inference]** Claus does have independent *assessment*. It lacks (a) independent *acceptance authority*, (b) a portable cross-swarm review record and (c) an IMD-style planner that splits one job across many seats. | | F2 | Cross-swarm review is new (phase 3) | **[Read]** Claus's [paid-deals guide](https://claus.si/docs/paid-deals) already resolves disputes through one IMD oracle request to a five-member panel. Four must give an identical payout (zero tolerance). A separate Claus decision follows, producing an expiring resolution signature that a party must submit. The [capabilities](https://claus.si/__network/v1/public/capabilities) endpoint advertised 7 remaining subsidised cases out of a finite 8. | V1, V2, V3, V4 (V5 notes the subsidised case) | **Corrected.** **[Inference]** An IMD–Claus review link already exists. Because it runs through IMD, it cannot serve as the *independent* arbiter for disputes about IMD's own work (V1–V4). | | F3 | Claus sells "x402 services in USDG" | **[Read]** Contract Control costs 0.25 USDC per address (0.50 USDC for an earlier-block comparison) and is paid by one-shot x402 in **Base USDC** ([customer services](https://claus.si/docs/customer-services), [products](https://claus.si/__network/v1/public/commerce/products)). Generic services use a separate direct-prepayment path on Sepolia ([services](https://claus.si/docs/services)). Jobs escrow takes ETH or CLAUS. Sponsored NFT inference settles in USDC on Base. USDG appears only in third-party City projects, for example Undertow at 1 USDG ([product reference](https://claus.si/product-reference.json), V5). | 5/5 | **Corrected.** Claus has several rails and no single asset. | | F4 | Claus has "no posted price" | **[Read]** Work profiles can quote prices in ETH or CLAUS ("offers, not funded agreements"), and commerce products have fixed USDC prices ([agent-economy](https://claus.si/docs/agent-economy)). | V1, V2, V4, V5 | **Narrowed.** **[Inference]** There is no posted price for a *coordinated upgrade review*. Prices do exist per agent and per product. | | F5 | The 2% pool fee is "half to buy and burn CLAUS, half to compute and hosting" | **[Read, conflict]** [migration-reference.json](https://claus.si/migration-reference.json) (`contentUpdatedAt 2026-11-06T21:23:59Z`) and the [migration guide](https://claus.si/docs/migration) say "1% buyback and burn plus 1% operating treasury". [llms.txt](https://claus.si/llms.txt) says "2% project fee, all allocated to liquidity… No current treasury share" and that buyback allocations are paused. *(spot-check: I confirmed both texts are live today.)* | V1, V2, V4 cite the 1/1 split (V2 and V4 flag the conflict). V5 cites only llms.txt and calls v0.1 wrong. | **[Uncertain]** The views disagree, and V5's "factually wrong" goes further than the evidence. Either way, the treasury half is described as an *operating treasury*, not "compute and hosting", and neither source makes a pool fee a service budget (V1–V5 agree). **What would settle it:** Claus states which document is authoritative, or someone reads the keeper's fee-routing transactions on Robinhood Chain. | | F6 | "New project inference submissions switched off on Oct 21 2026" | **[Read, conflict]** The [inference guide](https://claus.si/docs/inference) says "New project inference requests are disabled as of 21 October 2026". The [capabilities](https://claus.si/__network/v1/public/capabilities) endpoint returns `inference.enabled: true`, and the product reference says "enabled". *(spot-check: both confirmed today.)* | V1, V2, V3, V4 flag the conflict. V5 accepted "disabled". | **[Uncertain]** Neither reading should be a pilot dependency. Settled by the maintainers reconciling the docs. No one should test it by submitting requests. | | F7 | "395 API operations" | **[Read]** OpenAPI `0.41.0` counts to 426 operations. The HTML [API reference](https://claus.si/docs/api-reference) says 395. [Release status](https://claus.si/docs/release-status) (26 Oct) says 351. | V2, V3, V4, V5 | **Corrected.** **[Proposal]** Pin a schema version and digest. Do not quote a count. | | F8 | Claus has "no coordinator" | **[Read]** Claus has `workflow_create`, milestone briefs, managed workers and durable runtimes, plus a `coordinator` team-role preset that has no financial permissions. The [Build guide](https://claus.si/build.txt) separates NFT inference from an always-running agent. | 5/5 | **Narrowed.** **[Inference]** Claus has coordination, but no job-splitting control plane like IMD's run by house holders. | | F9 | The 300 houses can become seats | **[Read]** The [NFT guide](https://claus.si/docs/nft-inference) offers sponsorship of up to $5 per NFT per day, with automatic top-up off. "A key does not start an always-running agent." A transfer does not move private projects, history or keys. [NFT capabilities](https://claus.si/__network/v1/nft/capabilities) report `supports300Independent: false`. The sponsored model is `claude-haiku-5.5` with 1,024 max output tokens, 6 requests per minute and concurrency 1 (V5). Each NFT "represents 50,000 deposited $CLAUS; the current owner can burn it to redeem", and the NFTs remain on Ethereum ([llms.txt](https://claus.si/llms.txt), V5). | V1–V4 (property/transfer), V4 and V5 (`supports300Independent`), V5 only (model limits, burn-to-redeem) | **Corrected.** **[Inference]** Houses are property and access rights, not proven capacity. A seat tied to an NFT can disappear when the NFT is burned. Sponsored inference is too small a budget for upgrade review. | | F10 | CLAUS settlement is on Robinhood Chain | **[Read]** Jobs deal preparation accepts chain 1 or 11155111 only. The enabled [CLAUS escrow](https://claus.si/__network/v1/public/chains/1/deals?asset=CLAUS) uses the retired **Ethereum source token**. The current CLAUS is on Robinhood ([migration](https://claus.si/docs/migration), [home](https://claus.si/)). | V1, V3, V4 (V5: NFTs and stakes on Ethereum) | **Corrected.** A token symbol does not identify a settlement asset. | | F11 | ERC-8004 as a shared identity base | **[Read]** [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004) is a **Draft**. Payments are "orthogonal" to it. It acknowledges Sybil risk. IMD lists ERC-8004 Identity and Reputation registries ("Not deployed by IMD") on Ethereum mainnet and Sepolia only, none on Robinhood, and no Validation Registry ([IMD docs](https://imd.fun/docs/llms.txt), V5). Claus's public pages do not mention ERC-8004 (V5 search). | 5/5 (Draft), V5 only (deployment chains) | **Added fact.** **[Inference]** Records would sit on Ethereum while payments run on Robinhood (IMD) and Base (Claus). | | F12 | Huang's post is dated "Nov 12 2026" | **[Read]** V1 retrieved the [original post](https://x.com/matthuang/status/2120209025099043286), which shows **November 10, 2026**. V2, V3 and V4 did not verify it. V5's web search found no text using the terms. | V1 (found it), V5 (not found) | **[Uncertain, leaning corrected]** One view has a direct link. Whatever the date, "the rider owns it, no staff" is v0.1's extrapolation, not the post's requirement (V1). No design decision depends on it. | | F13 | CLAUS governance (not in v0.1, but it matters for job #1) | **[Read]** "The Robinhood token, hook and liquidity management are owner-upgradeable… single admin wallet with extra mint authority and no onchain mint cap or timelock… at least 1 hour of notice… not enforced by the contracts" ([llms.txt](https://claus.si/llms.txt), V5). The [migration reference](https://claus.si/migration-reference.json) also discloses a single developer admin with no enforced timelock (V2). | V5, V2 (V3 notes the root admin caveat in [coin reviews](https://claus.si/docs/coin-reviews)) | **Added fact.** **[Inference]** A review of an upgrade is advisory unless something onchain makes the admin wait for it. | **Confirmed as stated in v0.1 [Read]:** - IMD charges a flat 0.5 IMD per request, sold as admission, with no refunds and no guaranteed result. - IMD seats run on the holder's own machine and AI plan, and there are 2,000 of them. - 8% of each launched token goes to connected seats and 2% to wallets with accepted work. - IMD takes payment by x402 with Permit2, or through an intake contract with a callback. The Robinhood intake is `0xa43e6f75ee006411f79ac1c84120606c2330de82` on chain 4663, and the IMD token is `0x5f7b…7127`. - Claus job claims last one hour and only the requester accepts. - Claus mandates expire within 30 days. Sources: [IMD skill](https://imd.fun/skill.md), [IMD docs](https://imd.fun/docs/), [IMD onchain](https://imd.fun/skill/onchain.md), [IMD capabilities](https://api.imd.fun/requests/capabilities), [Jobs](https://claus.si/docs/jobs), [Build](https://claus.si/build.txt). Agreed by 5/5. **[Read] One correction about IMD that the design's bridge depends on.** IMD's HTTP route needs a second signature, an EIP-712 quote approval, on top of the x402 Permit2 signature. A stock x402 client does not produce it, and Safe-style contract wallets are not yet supported on that route ([pay.md](https://imd.fun/skill/pay.md)). Ordinary IMD job results are "Not signed by the oracle… the writer's word". Callbacks are made once with bounded gas, can arrive late or not at all, and failure callbacks are unsigned. The payer is the `msg.sender` of `request`, not `tx.origin`, and continuation rights stay with that payer ([onchain.md](https://imd.fun/skill/onchain.md)). Views: V1, V2, V3, V4, V5. The IMD workflow schema is strict, already publishes to IPFS/ENS, and documents no paid Claus step ([workflows.md](https://imd.fun/skill/workflows.md); V2, V4, V5). --- ## 1. The five changes, ranked Ranked by strength of evidence first, then by how many views propose the change. "Strength" means how directly first-party sources show that v0.1 as written would fail. ### Change 1: Keep assessment, acceptance, settlement and upgrade execution as four separate powers, built on Claus's existing review machinery **Proposed by 5/5 (V1 #2, V2 #2, V3 #2, V4 #2, V5 §1.4 / change 5).** Evidence: strongest. Claus's docs, OpenAPI and pinned code all state it directly. - **[Proposal] What.** Do not build a new acceptance system. Use `reviewer_candidates`, a `reviewPolicy` committed before the claim, frozen evidence, `app_task_review_submit` reports and a separate commission (V1–V5). For an upgrade, freeze the following and require a fresh review if any of it changes (V1, V2, V3, V4): - target chain, block and state; - current proxy, implementation and admin; - source commit and reproducible build settings; - old and new bytecode hashes; - storage layout; - exact upgrade calldata and expected postconditions. Pay the reviewer whatever the verdict, and keep `insufficient_evidence` as a valid result (5/5). The requester (`app_task_accept`) or the requester's explicit policy accepts. Only the project's existing authorised signer executes. A report never moves money or installs code (5/5). - **[Proposal] Replace "acceptance by an independent second run".** A second run is one method, not a definition of independence. The same controller, model or toolchain can repeat the same blind spot (V1, V2, V3, V4). V5 would still require a second run by a second seat with no shared-ownership flag, and would treat disagreement between the runs as `insufficient_evidence`. *Both positions hold:* the second run is useful evidence, provided its operator is different and conflicts are disclosed (V3: "preferably different model/tooling"). - **[Proposal] Replace "work above a set value gets the other swarm's review" with a risk class.** Any upgrade that changes authority, custody or issuance always gets the stronger review class, whatever the order is worth (V2, V3, V4; V1 cuts automatic mutual-review loops). V5 keeps a value threshold for routing the second review to IMD. *Disagreement:* value-based (V5) versus risk-based (V2, V3, V4). **[Inference]** The evidence favours risk: a cheap order can touch a valuable vault, and the CLAUS admin can mint without limit (F13). A value threshold could still sit *on top of* the risk class. - **[Proposal] Conflict-aware arbitration.** The existing dispute path buys an IMD panel. So for any job involving IMD, name a different, pre-agreed arbiter before funding, or decline the job (V1, V2, V3, V4). Treat the 8-case subsidy as finite (V1, V2, V4). - **[Proposal] Fit review into the escrow clock.** **[Read]** Delivery is followed by 24 hours to accept or dispute. Silence pays the provider. Disputes have 3 days, after which an unresolved dispute refunds the buyer. Arbitration admission needs ≥3 hours remaining, and objections do not stop the clock ([paid-deals](https://claus.si/docs/paid-deals), pinned [escrow](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/contracts/src/ClausDealEscrow.sol); V1, V2, V3, V4). Admit a review only if its turnaround fits inside that window. Label a timeout settlement as a timeout, never as "independently passed" (V1, V2, V4). - **[Proposal] Make the review bind only where a contract enforces it.** Proposed as a ranked change by V5 only (its change 5). Raised as a precondition or founder question by V2 and V4, and noted by V3. The point rests on F13 and is well sourced, so it is kept. Put the CLAUS Robinhood token, hook and liquidity owner behind an onchain timelock at least as long as the announced notice. Run the review inside that window, and cite the report hash in the `claus-journal/1` record (V5). V4: reject "review every upgrade" as a *safety guarantee* until administrators are bound by a gate. V2: a binding gate belongs in a separately reviewed execution module. **[Inference]** One hour is too short for a two-run review plus an IMD cross-review, since IMD results take "minutes to days" (V5). - **Cost.** - Reviewer fees that are owed regardless of outcome. - Storage for frozen evidence. - Conflict screening. - A deadline monitor run by the buyer. - Fewer jobs admitted when no independent reviewer is available. - A timelock means the Claus founder gives up instant upgrade and mint power, and the timelock needs its own audit (V5). ### Change 2: Price each bounded service in the asset it actually settles in; "one payment" becomes one brief plus an itemised authorisation **Proposed by 5/5 (V1 #3, V2 #5, V3 #4, V4 #4, V5 #3).** Evidence: strong. F3, F5 and F10 and the IMD second-signature fact are first-party. - **[Proposal] What.** - Drop the single posted CLAUS-over-x402 price. - Each service advertises its exact chain, asset contract and decimals, atomic amount, payee, mechanism, expiry, and what the payment buys: admission, a snapshot, effort, or an accepted outcome (5/5). - Price upgrade review in bounded classes (chain, contract count, diff size, privileges and value exposed), with explicit exclusions (V1, V2, V3, V4). - CLAUS on Robinhood becomes an *additional* rail once Claus actually runs it (V1, V5). - Decouple the burn/seat/treasury split from pool fees. Publish it per service, and only once there is revenue (V2, V4, V5). - **[Proposal] One payment.** - Pilot: one brief, an itemised budget authorisation, and native child payments (V1, V2, V3, V4). - A literal single payment requires an explicit merchant or router. That party holds inventory, quotes FX, slippage and gas, and states who absorbs a failed child order. No one can refund IMD admission that has already been spent (5/5). - V5 states the same idea as "the rider pays the *first* swarm, which subcontracts", with the first swarm taking the currency risk. **[Inference]** That is the merchant model with the merchant named. It is consistent with the others, as long as it is disclosed. - **Why [Inference].** x402 negotiates a payment. It does not make two services atomic or arrange token conversion ([x402 networks](https://docs.x402.org/core-concepts/network-and-token-support); V1, V3). A stock x402 payer cannot buy from IMD's HTTP route (see the IMD correction at the end of §0). - **Cost.** - A weaker "every job burns CLAUS" story. - Quote maintenance. - Claus must hold IMD on Robinhood to order IMD jobs (V2, V5). - Any merchant carries custody, accounting and loss risk. ### Change 3: Replace the "bridge both ways" and the six-question card with a versioned work-agreement profile, signed receipts and a machine-readable card **Proposed by 5/5 (V1 #1, V2 #1, V3 #1 and #5, V4 #1, V5 #1).** Evidence: strong, but the need is *inferred*. The sources show that existing standards each cover only part of the problem. A2A messages "MAY be idempotent". ERC-8004 leaves payments out. x402 receipts carry no rubric or output digest ([A2A](https://a2a-protocol.org/latest/specification/), [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004), [x402 offer/receipt](https://docs.x402.org/extensions/offer-receipt)). Native IMD and Claus semantics already separate the states that v0.1's word "shipped" merges. - **[Proposal] Smallest interface (agreed in substance by 5/5).** Five operations, each mapped onto existing native endpoints through one adapter per side. No new universal API. | Operation | Must bind | Already exists natively | |---|---|---| | Discover | Signed, versioned card: accountable provider, service/schema versions, endpoints, keys and their powers, payment profiles, review and failure policy, expiry | A2A card shape. No swarm card yet. | | Quote | Buyer, provider, service version, input digest, output schema, chain/asset/amount/payee, deadlines, acceptance/dispute/refund rules, onward-disclosure permission, expiry | IMD `/requests/check` → `/requests/quote` (10 min). Claus `service_quote`/`service_accept` (V5). | | Order | Buyer-generated idempotency key, quote hash, payment reference, native order ID, parent/root job link | IMD intake or x402. Claus escrow, x402 or direct prepayment. | | Read | Separate states for execution, review, acceptance and money, including `unknown`. A read never triggers a second payment. | IMD `/intake/:chain/:tx`, `/jobs/:id`. Claus `service_order`, `public_review`. | | Record | Delivery receipt, review receipt(s) and buyer decision as distinct signed receipts bound to exact artifact and rubric hashes, with conflicts and supersession | Claus review reports. IMD `acceptance-v2`/WorkRegistry. | - **[Proposal] Recovery rules (5/5).** - The same key with the same bytes returns the original order. The same key with different bytes is a conflict. - Persist the intended order and the transaction nonce or hash *before* signing. - An uncertain payment stays `unknown` and is reconciled. It never authorises a replacement purchase. - Polling is mandatory. Webhooks are optional hints. - The x402 payment-identifier cache has a TTL, so a durable application ledger is still needed ([payment-identifier](https://docs.x402.org/extensions/payment-identifier); V1, V3, V4). IMD's strict parser rejects non-empty payment extensions today, so IMD's own request key has to provide recovery (V4). - **[Proposal] Receipt encoding (disagreement on detail).** - V5: EIP-712-signed `{requestId, resultHash, uri, signer, signature}`, signed by a key named in the card. This requires IMD to begin signing job receipts it currently leaves unsigned. - V1: RFC 8785 canonical JSON, a SHA-256 document digest and a documented signature envelope. V1 and V4 keep native IMD keccak-256 and Claus SHA-256 commitments as tagged fields, not rehashed. - **[Inference]** These are compatible: EIP-712 for onchain-verifiable signatures, with both native commitments carried inside. Settled by whichever format a third swarm can verify without trusting either API. Test it in phase 3. - **[Proposal] The card.** The six questions are the right *human* headings (5/5), but they do not work as a protocol (5/5). Every view adds an authority-and-failure field: - V1: "Who can authorize, change, revoke or dispute this service?" - V2: "Where is the versioned machine interface?" and "Who can authorize what, and what happens when it fails?" - V3 and V4: "Who may act, and what happens on failure?" - V5: identity anchor, endpoints and schemas, keys and powers, failure terms, independence policy. Further changes: - Reword "who are your seats" as "who is accountable, and how is membership attested". A seat roster becomes optional (V1, V2, V3, V4). V5 wants it checkable onchain through a registry or adapter. - "Who checked it" is a fact about each result. The card states the *policy* (V2, V4). - The cost field points to an expiring quote endpoint (V1, V2, V3). - Bind each order to the card and quote versions in force when it was placed (V3, V4). - **Disagreement on card format.** V5 wants an ERC-8004-style registration file for the swarm at `/.well-known/swarm-card.json`, linked onchain. V1 and V3 want an A2A-compatible card with a named commercial extension, claiming A2A only where an A2A endpoint really exists. V2 makes A2A optional. **[Inference]** Both can be one document: ERC-8004 registration files already carry `services` entries that can point to an A2A card. Settle it by having the third swarm's adapter consume only the published file. - **Cost.** A short spec, two adapters, durable deduplication and order journals, conformance fixtures, and a guarded, rotatable receipt key per swarm. The rider dashboard comes later. ### Change 4: Keep house (property), agent identity, operating keys and rider record separate; use ERC-8004 only for accountable, persistent identities **Proposed by 5/5 on the principle (V1 #4, V2 #4, V3 #3, V4 #3, V5 #2 and #4).** They disagree on whether each house gets an ERC-8004 identity. Evidence: strong for the facts (ERC-8004 Draft semantics, NFT transfer rules, `supports300Independent: false`, burn-to-redeem). What to register is a proposal. - **[Read]** The ERC-8004 global identifier is registry namespace/chain/address plus `agentId`, an ERC-721 resolving to a registration file. `agentWallet` is cleared on transfer and must be re-proved. Clients may give feedback without registering. `getSummary` requires a `clientAddresses` filter because unfiltered results "are subject to Sybil/spam attacks" ([ERC-8004](https://eips.ethereum.org/EIPS/eip-8004); 5/5). IMD binds seat NFT → device → agent ID through a separate adapter and writes verdicts through WorkRegistry ([IMD API](https://imd.fun/docs/api/); V1, V2, V5). - **[Proposal] Agreed by all five:** - Riders are *not* registered as agents. A rider is a client principal (wallet or organisation account plus signed delegations). A rider's record is an opt-in set of signed receipts that both swarms index. There are not two independently edited scores (5/5). A rider who runs an agent service may register *that service* (V1, V3, V4). - Identity owner, payment wallet, API credential and upgrade admin stay distinct (5/5). - When control changes, history stays with the operator and the ownership interval in which the work was done. A buyer of an identity does not inherit skill (V1, V2, V3, V4). On burn or redemption the seat retires and its history stays (V5). - Register no fictional workers to make up a count (V2, V4). - **Disagreement: what to register for houses.** - *V1, V2, V3, V4:* register the Claus gateway first, then individual persistent workers that actually do attributable work. A house is a linked membership/property claim (chain/collection/token/interval), not the ERC-8004 locator. - *V5:* a house becomes a seat only after the holder registers an ERC-8004 ID through a Claus adapter (the same pattern as IMD's), runs a worker on their *own* plan, and passes a paid trial review. - **[Inference]** These converge more than they appear to. V5's gating (own plan, trial review, burn retires) produces exactly the "persistent workers that do attributable work" the others would register. The live question is whether a house is the *only* route to a Claus worker identity. **Settled by** founder question O4 (does a house confer a right to work, an obligation, or only property and sponsorship?). - **Disagreement: reviews onchain now or later.** - *V5:* each swarm writes verdicts from one contract (IMD WorkRegistry and a new Claus ReviewRegistry). Cross-swarm reputation is then `getSummary(agentId, [IMD WorkRegistry, Claus ReviewRegistry])`. The independent second run uses the Validation Registry with the other swarm as validator. That would need a Validation Registry deployment, which no one lists. - *V1:* signed offchain receipts first, with issuer-attributed batches anchored onchain later. - *V2, V3, V4:* export selected, evidence-linked feedback. Use validation only where a real validator mechanism is deployed. - **[Inference]** V5's client-filter design is precise and cheap to state, and it matches how IMD already works. The cost and dependency are a new Claus contract plus audit, Ethereum gas, and a Draft standard with no Validation Registry deployed. **Settled by** IMD agreeing to treat a Claus review contract as a trusted `clientAddress` (O9) and an agreed Validation Registry address (O10). Until then, offchain signed receipts lose nothing. - **Cost.** Identity mapping, ownership-interval indexing, key rotation and revocation, registry gas, privacy choices, and no simple "300 houses = 300 workers" leaderboard. ### Change 5: Gate rollout on demonstrated interoperability and recovery, not on seat counts or a company launch **Proposed by 5/5 in some form (V1 #5, V4 #5, V3 #5 pilot, V2 phases and cuts, V5 #4 "honest capacity" and its phase gates).** Evidence: moderate. `supports300Independent: false`, the finite arbitration subsidy and the inference conflict are read from sources. The gate design is a proposal. - **[Proposal] What.** - Keep Phase 0 and Phase 1 as they are. - Before scaling, run failure drills: lost response, duplicate submit, expired quote, unknown payment, changed artifact, offline reviewer, revoked credential, callback loss, deadline or dispute fallback (V1, V2, V3, V4). - Seed known defects to measure what reviewers catch (V1, V3). - Treat 30 as a *ceiling* of voluntary, qualified operators, not a target (V1, V2, V4). - Admit a third, independently run swarm adapter *before* declaring the interface stable. It must work from the published schemas with no bilateral fork (V1, V2, V3, V4). V5 opens the card to a third swarm in Phase 4. - Publish quote-to-result latency, duplicate charges, unresolved orders, reviewer disagreement and cost per accepted result (V4). - Measure "still live later" against a stated horizon. A failed check stays `unknown` (V2, V4). - **[Proposal] The swarm-native company moves later.** Before any launch, name an accountable owner for keys, contracts, hosting bills, incidents and recovery (V1, V2, V3, V4). V5 allows it in Phase 4 only if Phase 3 shows revenue, and raises the legal ownership question. - **Cost.** A smaller launch story, paid rehearsal work and a corpus for qualifying reviewers. --- ## 2. What I would cut | Cut | Views | Basis | |---|---|---| | Mandatory conversion of all 300 houses to holder-funded seats as a prerequisite, and any mandatory public seat roster | V1, V2, V3, V4. V5 keeps holder's-own-plan as *the* seat runtime but admits seats one at a time. | F9 **[Read]**. A membership policy is not an interop requirement **[Inference]**. | | One flat CLAUS price for every request, and any link between pool-fee burn and service affordability | 5/5 | F3, F5, F10 | | "Half to burn, half to compute and hosting" as a design input | 5/5 (V5: "factually wrong"; others: a trading-fee policy, not a service budget) | F5 (conflict noted) | | A value-only threshold for cross-swarm review, and automatic mutual-review loops | V1, V2, V3, V4 (V5 keeps the threshold) | F13, F2 | | Rider priority, credit and fee share in the pilot. Keep the receipt record. | 5/5 | ERC-8004 Sybil warning; no revenue to share (V5) | | A second, independently edited rider database | V1, V3, V4 (V2, V5 equivalent) | Inference | | The word "bridge", wherever it implies asset bridging or shared signing | V1, V3, V5 (Claus migrated with "No CCIP", [llms.txt](https://claus.si/llms.txt)), V2 | **[Read]** / Proposal | | A paid Claus hosting step at the end of every IMD workflow | V4, V5 cut it, because IMD already publishes to IPFS/ENS. V1, V2, V3 keep three City publications but say free publication must not be presented as paid hosting. | [workflows.md](https://imd.fun/skill/workflows.md) | | Automatic upgrade execution when a review completes | 5/5 | F1 | | "No staff" as a launch promise, and the first company as a scale gate | 5/5 (V5: move to Phase 4 conditional on revenue) | Inference | | Operation counts and other fixed numbers copied from prose into the design | V3, V4, V5 | F7 | ## 3. What I would keep exactly as is - **"Adjacent and not a copy"**, and **the link must work for a third swarm**: 5/5. - **"No shared token. Each swarm is paid in its own."** 5/5. Read this as "each *service* declares the asset it settles in", per Change 2. - **Upgrade review as the first coordinated job type**: 5/5. V5 adds that the governance facts (F13) make it the most valuable job available. - **Phase 0, research and a yes or no**: 5/5. **Phase 1, three IMD jobs ordered by Claus and three IMD builds published in the City**: V1–V4 keep it exactly. V5 keeps "hire each other with what exists" but swaps in three Claus reviews ordered by IMD over x402. - **The six questions, verbatim**, as the card's human front page: 5/5. - **One brief** as the rider's entry point: V1–V4. Keep it; the one-*payment* promise changes. - **The IMD builds / Claus houses, operates and reviews division**: V2, V5. V2 notes the roles overlap and should not be exclusive. - **Seats on the holder's own machine and plan**: V5 only, as a hard requirement for any Claus seat. The others leave the funding source to Claus's internal policy, as long as it is disclosed per job. - **A 30-seat pilot before 300**: V3 (V1, V2 and V4 keep 30 only as a ceiling). --- ## 4. Calls between the swarms and what each side must trust **Claus app or hook → IMD job (5/5 agree).** - **[Read]** `requestWithFailure(bytes32,bytes,(address,bytes4),bytes4,address,uint256)` on the Robinhood intake, 0.5 IMD, public JSON body ([onchain.md](https://imd.fun/skill/onchain.md)). - **[Proposal]** Preflight the body offchain. Approve the exact amount from a bounded signer or a purpose-built ordering contract. Pin the intake, action version and maximum price. Persist the request ID and transaction. Set the launch `owner` and allocation recipients explicitly, because the payer contract otherwise becomes the default. - **[Proposal]** A hook records a work *intent*. A keeper or owner-funded adapter submits it outside the swap, with deduplication, rate limits and a pause. - **[Proposal]** The callback receiver checks the intake sender and its own pending request ID, stores the result, and does nothing consequential in that callback. - **[Proposal]** If a hook needs a yes or no, ask the IMD oracle a bound question and verify `questionHash`, consumer, quorum and freshness (V5 and V1–V4). Never act on an unsigned job callback. - **[Read]** The Claus app-service sandbox has no network, payment or secret access ([OpenAPI](https://claus.si/__network/openapi.json); V1, V4). So an external adapter or the intake contract is required. - **Trust.** Claus trusts IMD's writer key for result attribution only: it "cannot forge an answer but can withhold one". It trusts the oracle signer for bound answers. IMD trusts nothing about Claus keys, because payment is pulled in the same call (V5). **IMD workflow → paid Claus service (disagreement).** - *V1, V2, V3, V4:* an external adapter, outside IMD workers, runs after the IMD result. It pins the artifact manifest, gets a Claus quote or a funded Jobs agreement, and pays from its own budget-capped signer. It holds a narrowly scoped Claus project credential where publication is needed. - *V5:* pay a Claus service endpoint directly over x402 and give IMD no Claus credential at all. Mandates expire within 30 days, and `service_quote` needs `service.call`. IMD then holds only a capped Base wallet. - *Both agree:* - The first paid exercise is Contract Control (`GET /v1/x402/contract-control`, Base USDC), labelled as an observation and not an audit (V1, V2, V4, V5). - A paid upgrade-review or hosting endpoint does not exist yet and must be published first (V1–V5). - A native IMD workflow step would be new IMD work, because the workflow schema is strict (V2, V4). - **[Inference]** Settled by product type. Review and observation services can be pure x402 with no shared credential (V5's point). City publication needs project authority (V1–V4's point), so it should use a scoped, short-lived credential held by the adapter, never by IMD workers. **Trust rules (5/5).** - Neither swarm holds the other's owner, treasury, upgrade, seat-device or model-provider keys. - Keep separate keys for receipt signing, quote signing, payment, API delegation, review, arbitration and upgrades. Use the x402 offer/receipt extension's separate signing key ([offer-receipt](https://docs.x402.org/extensions/offer-receipt)). - Delegations bind audience, operation, agreement, spend cap, nonce and expiry. - Keep authorisation history so old receipts still verify after key rotation. Revocation does not undo transfers already confirmed. - A compromised connector loses only its budget. - Do not pass bearer tokens through MCP ([MCP security](https://modelcontextprotocol.io/docs/2025-11-25/tutorials/security/security_best_practices); V3, V4). **Other systems (5/5 on the pattern).** Each has something to borrow and something to avoid: - **A2A:** borrow cards, task IDs, artifacts and polling. Do not treat `completed` as buyer acceptance or payment. Pin a version. - **x402:** borrow expiring offers, separate receipt keys and payment identifiers. Do not claim exactly-once delivery or refunds from it. - **ERC-8004:** borrow registry-scoped locators and client-filtered feedback. Do not treat a registration as authority or independence. - **Virtuals ACP / ERC-8183:** borrow the client, provider and evaluator roles. In its quickstart the buyer can be the evaluator, so a role name does not prove independence (V1, V2, V4). - **MCP:** use it as a tool surface, not as consent to spend. - **AutoGen and OpenAI Swarm handoffs:** borrow bounded handoffs. In-process routing is not a commercial contract (V2, V3). - **Anthropic's multi-agent research system:** borrow precise briefs and outcome evaluations, and budget for the token overhead (V1). --- ## 5. Open questions only the founders can answer | # | Question | Raised by | |---|---|---| | O1 | Who is the seller accountable to the rider for a combined order, and who absorbs a spent IMD admission fee when the Claus leg fails? Does "one payment" mean one approval, one invoice, one transfer, or atomic settlement? | V1, V2, V3, V4, V5 | | O2 | Will Claus put the Robinhood CLAUS token, hook and liquidity owner behind an enforced onchain timelock or review gate? If not, is an advisory review worth paying for? Who holds emergency override? | V5, V2, V3, V4 | | O3 | Which chain and asset will new Claus review payments use: CLAUS on Robinhood, USDC on Base, or both? What happens to the Ethereum source-token escrow? Will Claus hold IMD on Robinhood? | V1, V2, V3, V4, V5 | | O4 | Does a house confer a right to work, an obligation to operate, or only property and sponsorship? Is a house the only route to a Claus worker identity? What happens to open jobs and history on transfer or on burn-to-redeem? Does sponsorship continue for seated houses? | V1, V2, V3, V4, V5 | | O5 | Which affiliations disqualify a reviewer or arbiter (shared controller, funder, model provider, subcontracting)? Who arbitrates disputes where IMD or Claus is a party? Who pays for that, and can they meet escrow deadlines? | V1, V2, V3, V4 | | O6 | Is independent review advisory, required by a requester policy, or a binding condition on execution? Which risk classes require external review, and is there also a value threshold? Who pays when the two swarms' verdicts disagree? | V2, V3, V4, V5 | | O7 | What does "the rider owns the company" convey: admin roles, repository and domain, legal entity, revenue? Who is liable and on call when there are "no staff"? | V1, V2, V3, V4, V5 | | O8 | What is public in rider records? Can a rider opt out? What evidence would later justify priority, credit or fee share, and who funds the losses? | V1, V2, V3, V4 | | O9 | Will IMD sign job receipts with a key published in its card? Will IMD treat a Claus review contract as a trusted `clientAddress` and a Claus verdict as an independent second review? | V5 | | O10 | Which ERC-8004 registry deployments and IDs will each team endorse, and who deploys or chooses a Validation Registry? | V4, V5 | | O11 | What budget, per-order and total loss ceilings, and reliability and margin targets must be agreed *before* the first pilot purchase? | V2, V4 | **Not founder choices: publication fixes the maintainers owe [Uncertain].** - The fee-allocation conflict (F5). - The inference flag conflict (F6). - The stale operation counts (F7). - The date of the Huang post (F12). The proposer should link the post or present the thesis as their own (V5). --- ## 6. Revised design v0.2: one page (proposal; nothing agreed by either team) **Goal.** Claus is an independent swarm beside IMD, adjacent and not a copy. IMD mainly sells builds and launches. Claus mainly sells bounded upgrade review, City publication and operation. The roles overlap and are not exclusive. A rider sends one brief, approves an itemised, bounded plan, and gets an attributable record. A third swarm joins through the same profile, with no bilateral code. There is no shared token, scheduler, treasury, bridge or set of keys. **Corrected facts it rests on.** - Claus already has precommitted, evidence-frozen, advisory task review, reviewer matching, commissions and escrow arbitration, and that arbitration runs through an IMD panel. - Claus settles in Base USDC (commerce), ETH or Ethereum-CLAUS (Jobs escrow) and Sepolia (generic services). It has no USDG and no Robinhood escrow. - The pool-fee allocation is published inconsistently, and it is not a service budget. - Houses are Ethereum property NFTs, burnable for 50,000 CLAUS, with small sponsored inference and no confirmed capacity for 300 at once. - The CLAUS Robinhood contracts are single-admin with no enforced timelock. - IMD sells 0.5-IMD admission with no refunds. Its job results are unsigned writer assertions and its oracle answers are signed. Its HTTP route needs a second EIP-712 signature. - ERC-8004 is a Draft, with registries listed on Ethereum only. **Protocol (what a third swarm adopts).** 1. *Card:* one signed, versioned file at a well-known URL. It carries the six questions verbatim plus a seventh, "who may act, and what happens on failure". Machine fields: accountable identity (optional ERC-8004 locator), service/schema versions, endpoints, keys and their powers, quote endpoint, exact chain/asset/mechanism tuples, review and independence policy, dispute, refund and retention terms, capacity observation with expiry. Orders bind to the card and quote versions in force when placed. 2. *Five operations:* Discover, Quote, Order (idempotent), Read (separate execution, review, acceptance and money states, `unknown` allowed), Record (signed delivery, review and decision receipts bound to exact hashes; native keccak and SHA-256 commitments preserved). Polling is mandatory. An uncertain payment never triggers a repurchase. 3. *Calls:* Claus → IMD goes through the Robinhood intake from a bounded signer or ordering contract. Hooks record an intent and a keeper submits it, and callbacks only store results. IMD → Claus goes through an external adapter with its own capped wallet: pure x402 for review and observation, a scoped short-lived credential for publication. A first paid run uses Contract Control and is labelled an observation. 4. *Trust:* each side signs only its own claims and keeps separate receipt, payment, review, arbitration and upgrade keys. Value moves only on a bound oracle answer or an explicit signer decision. **Assurance.** Upgrade review covers a frozen bundle: source, build, old and new bytecode, storage layout, calldata, chain state and rubric. A preselected reviewer with conflicts disclosed reviews it. The reviewer is paid whatever the verdict, and `insufficient_evidence` is allowed. A second, independent operator reviews anything that changes authority, custody or issuance. The requester accepts, the escrow settles, and the owner executes; these stay separate. Review deadlines fit inside the 24-hour and 3-day escrow windows. Disputes involving IMD use a pre-agreed arbiter outside IMD. The review binds upgrades only once an onchain timelock of at least the notice period is in place (O2). **Identity and record.** ERC-8004 identities go to accountable gateways and to persistent, qualified workers. A house links to its worker through time-bounded operator mappings, and a seat retires on burn. Riders are client principals with opt-in signed receipts that both swarms index. Reviews go onchain under a client-filter design once O9 and O10 are answered. There is no combined score, and no credit or fee share yet. **Money.** Each service has a bounded, quoted price in an asset it actually settles. Child payments are native and itemised under one approved brief. A single invoice comes later, from a named merchant with disclosed inventory, FX and loss policy. Any burn/seat/treasury split applies to service revenue under published terms, never to pool fees. **Phases.** - **Phase 0:** reconcile F5 and F6, answer O1–O3 and O11, then decide yes or no. - **Phase 1:** three IMD jobs ordered by Claus and three IMD builds published in the City, with linked receipts. Free publication is labelled honestly. Add one real paid x402 call. - **Phase 2:** up to 30 voluntary, qualified operators on upgrade review. Seeded defects and the full set of failure drills. Behind a timelock if O2 is answered yes. - **Phase 3:** cards live on both domains, the receipt profile frozen, and a third independent adapter passing conformance without bilateral changes. Opt-in rider records. - **Phase 4:** grow toward 300 only on measured demand, quality, latency and cost. Rider incentives and a swarm-native company follow only after named owners, an incident owner, funded operations and exit paths exist (O7, O8). **Not in scope.** A shared token or bridge, a mandatory house conversion, paid hosting inside every IMD workflow, rewards funded from pool fees, and automatic upgrades triggered by a review.