# Claus beside IMD: change the contract between swarms first Researcher 4 · protocol and interface · public-source desk research · 12 November 2026 **[Proposal] Proceed with a bounded reciprocal hiring pilot, but reject v0.1 as a protocol specification.** Make the shared primitive a versioned work agreement with separately verifiable payment, delivery, review and acceptance receipts. Do not make conversion of 300 houses, compulsory CLAUS payments, a rider credit system or a staffless company prerequisites for hiring across networks. **[Read / method]** Sources below were retrieved through public website/documentation GETs, including public OpenAPI and capability responses. No login, private endpoint, transaction, signature, paid request or write endpoint was used. API descriptions are evidence of the published interface, not proof of successful execution. No other researcher's view was consulted. The report does not independently audit contracts, establish actual operator independence, or measure service availability. “Current” below means the retrieved observation, not a continuing guarantee. **[Proposal / notation]** Every substantive paragraph, list item or table row is labelled: **Read** means a source says it; **Inference** means my conclusion from cited evidence; **Proposal** means a design choice I recommend; **Open** means a question requiring the founders' decision. Read is attribution, not endorsement. Proposals and the revised design are not claims of team agreement. The supplied Huang thesis is treated as motivation; its exact post/date is not independently established here and is not evidence for a protocol requirement. ## 1. Correct the premises before changing the design | v0.1 premise | Evidence and correction | Design consequence | | --- | --- | --- | | Claus has “no independent acceptance or onchain record.” | **[Read]** Claus Jobs already supports a reviewer selected before claim, frozen input/output/evidence, an authenticated immutable report with criterion-level `pass`, `fail` or `insufficient_evidence`, objections and a separately paid review commission. Review is advisory: the requester still accepts. Paid Jobs bind terms and results to onchain escrow commitments. [Jobs][s1] [OpenAPI][s2] | **[Inference]** Independent review exists as an interface; independent final acceptance is not enforced by that review. The gap is policy and cross-network receipt portability, not absence of review or onchain commitment. | | Claus has no posted price and services use USDG. | **[Read]** Work profiles can publish ETH/CLAUS quotes. Customer Contract Control lists 0.25 USDC per address, or 0.50 USDC for an earlier-block comparison; its retrieved product catalog enables that offer. Its documented one-shot x402 route takes Base USDC, not USDG. Generic registered services have their own quoted direct-prepayment path. [Work profiles][s3] [Customer services][s4] [Products][s5] [Service protocol][s6] | **[Inference]** “No price for a new coordinated upgrade-review product” may be true, but “no posted price” is too broad. Do not replace distinct existing payment products with a presumed universal USDG or CLAUS rail. | | New project inference remains switched off after 21 October. | **[Read]** The services page still says new project inference is disabled. However, the retrieved public capabilities say `inference.enabled: true`, and the product reference reports project inference enabled. NFT sponsorship and native resident funding are separate products. [Services page][s7] [Capabilities][s8] [Product reference][s9] | **[Inference]** Treat the shutdown as a historical statement, with present documentation drift. Enabled capability is not proof an arbitrary project has an allowance or can execute. Runtime admission needs a fresh capability/allowance check. | | A 2% Robinhood pool fee is split equally between burn and treasury. | **[Read]** The migration reference, `contentUpdatedAt` 6 November, states 200 bps split into 100 bps buyback/burn and 100 bps treasury, with liquidity allocation stopped. `llms.txt` instead says all 2% goes to liquidity and no treasury. [Migration reference][s10] [Discovery text][s11] | **[Inference]** v0.1 agrees with the dated reference; it is not safely refuted by the conflicting discovery text. The live allocation remains unverified in this desk review. Neither a pool fee nor its allocation establishes the economics of a future service price. | | Claus has 395 API operations. | **[Read]** The retrieved OpenAPI declares version `0.41.0`; counting HTTP methods GET/POST/PUT/PATCH/DELETE gives **426** operations. The release page, dated 26 October, instead advertises 351. [OpenAPI][s2] [Release status][s12] | **[Inference]** Pin the schema/version used by an adapter; operation count is a poor readiness metric. | | Three hundred houses can become three hundred working seats. | **[Read]** NFT capabilities report admission enabled and provider validation, but explicitly `supports300Independent: false`. The product reference describes up to $5 per NFT per day and says simultaneous 300-request execution is not confirmed. [NFT capabilities][s13] [Product reference][s9] | **[Inference]** Houses, inference entitlements, independent operators and qualified review workers are different sets. A catalog size is not a capacity or independence demonstration. | | Both sides can use Robinhood as the common settlement home. | **[Read]** IMD advertises Robinhood intake. Claus's documented Jobs deal preparation accepts chain 1 or Sepolia, and its retrieved enabled CLAUS escrow uses the **Ethereum** token, not the Robinhood CLAUS token. [IMD intake][s14] [OpenAPI][s2] [CLAUS deal terms][s15] | **[Inference]** An IMD order on Robinhood is possible without implying that a linked Claus escrow can settle there. Every service needs an explicit chain/asset pair. | **[Read]** IMD's public docs substantiate the 2,000-seat, own-machine/own-plan model and the 8% connected-seat/2% accepted-work launch allocations. Its payment guide says 0.5 IMD purchases **action admission**, with no guaranteed result; the skill text warns that refused bodies can still consume payment. These are not outcome-contingent escrow terms. [IMD docs][s16] [IMD payments][s17] [IMD skill][s18] **[Inference]** “No coordinator” is also too absolute: Claus documents managed workers, service dispatch and coordinated review/arbitration flows. These sources do not establish a decentralized house-holder scheduler equivalent to IMD. Specify that narrower missing capability rather than equating it with an empty platform. [Release status][s12] [Service protocol][s6] [Paid deals][s19] ## 2. The five changes, ranked ### 1 — Replace the informal bridge with a work-agreement protocol and receipts **[Proposal / change]** Standardize only discovery, a pinned agreement, idempotent ordering, status/result retrieval and attributable decisions. Each swarm keeps its scheduler, seats, token economics and internal checks. The caller hires the receiving swarm as the accountable provider; internal subcontractors can be disclosed without becoming new counterparties automatically. **[Inference / reason]** An HTTP success, a paid transaction, a completed job and a safe upgrade are different claims. Claus already separates task acceptance from escrow acceptance and withdrawal; IMD distinguishes admission from completion and launch deployment. Combining those into “shipped” loses material information. [Jobs][s1] [IMD intake][s14] [IMD workflows][s20] **[Proposal / concrete change]** Use the minimal operations and receipt fields in section 3. Store a request commitment and native order mapping before authorizing payment; persist the original payment nonce/hash before broadcast. Same requester, idempotency key and input commitment must recover the same order; changed input must conflict. Return explicit `unknown` payment state after an uncertain broadcast. A workflow resumes that order, rather than buying a replacement. **[Proposal / cost]** Two adapters, a canonical serialization profile, a durable order/receipt store and contract tests for duplicate requests, lost responses, conflicts and partial completion. This is more work than six descriptive answers, but far less than a shared scheduler or a new settlement chain. The interface adds no mandatory joint token or joint governance contract. ### 2 — Make acceptance a precommitted policy; reuse Claus reviews and arbitration **[Proposal / change]** Commit the reviewer, observable rubric, evidence scope, deadline, payer and decision authority before work starts. Pay review work regardless of favorable or unfavorable findings. Separate four decisions: provider delivery, independent technical assessment, buyer acceptance and financial settlement. A second model run is one piece of evidence, never a sufficient definition of independence. **[Read / reason]** Claus already implements this distinction for Jobs. Its paid-deal dispute path separately uses an exact five-member IMD oracle question, at least four agreeing with zero numerical tolerance, followed by a genuine Claus decision on that payout split. A returned resolution signature still needs a party's transaction. The retrieved capability allocation has seven remaining cases from a finite eight-case subsidy. [Jobs][s1] [Paid deals][s19] [Capabilities][s8] **[Inference / reason]** “Get review from the other swarm” does not automatically supply an independent adjudicator: a Claus-versus-IMD job could otherwise route its dispute back to IMD. Different token IDs, model instances or swarm names do not establish separate controllers. [Paid deals][s19] [Work profiles][s3] **[Proposal / concrete change]** Start with the existing advisory report plus explicit buyer signing. For contested funds, retain the actual escrow rules and identify their arbiter in the agreement. Exclude economically involved builders, operators and original reviewers from dispute adjudication. If the existing arbitration lane cannot satisfy that conflict rule, disclose it as unavailable for that job and select another mutually agreed arbiter before funding—or do not admit the job. Do not promise a new arbiter for an already immutable agreement. **[Proposal / cost]** Reviewer fees, scheduling and conflict disclosures; a deadline monitor and an evidence store; lower throughput when a genuinely separate reviewer is unavailable. Requiring binding independent acceptance later costs a separately reviewed escrow or authorization-policy change. It cannot be accomplished by relabelling today's advisory report. ### 3 — Separate identity, membership, authority and reputation **[Proposal / change]** Keep houses as access/property identifiers. Register persistent service/agent identities under ERC-8004 only where that helps discovery and feedback. Publish a time-bounded mapping from house/NFT to operator, worker identity and authorized signing keys. Do not automatically mint 300 agent identities or require every rider to become an ERC-8004 agent. **[Read / reason]** ERC-8004 is Draft. Its global agent locator is registry namespace, chain, registry address and agent ID. The identity is ERC-721-based and transferable; it can link A2A/MCP endpoints. Payment is outside the standard. `agentWallet` must be reverified after transfer. Feedback excludes the owner/approved operator but does not prove that other reviewers have independent controllers. [ERC-8004, identity and security sections][s21] **[Proposal / concrete change]** Register a Claus hiring gateway and an IMD hiring gateway if the teams choose, then individually register enduring review workers that actually need portable reputations. Register neither “house number = proven reviewer” nor “rider score = trustworthy borrower.” Use the rider's authenticated principal and signed delegation as the initial identity; if a rider runs an agent service, that service may separately register under ERC-8004. Keep operator-era history across NFT transfers so a new controller does not silently inherit an old controller's qualification. **[Proposal / registration detail]** Each registered gateway or worker gets its own `(agentRegistry, agentId)` and `agentURI` resolving to registration metadata: name/description, advertised service endpoints and versions, active status, registry references and supported trust models. Bind the payment wallet using the registry mechanism, but publish delegated operational keys and permissions separately. The house collection and its token ID are a linked membership claim, not the ERC-8004 locator. The gateway can describe a swarm without registering every ephemeral model process. [ERC-8004][s21] **[Proposal / cost]** Registry transactions, metadata hosting, key rotation/revocation and operator-era bookkeeping. Publishing controller relationships exposes some commercial information; support explicit unknown relationships instead of pretending a complete Sybil detector exists. This preserves access for a third swarm that has no seat NFTs. ### 4 — Replace “one CLAUS price” with typed quotes and an explicit payment adapter **[Proposal / change]** Keep posted prices **per bounded service/version**, not one flat price for arbitrary upgrade reviews. Quote exact chain, asset contract, atomic amount, recipient, expiry, payment mechanism, purchased obligation, refund policy, review cost and any split. Keep pool-fee economics out of this interface. Native token preference is commercial policy, not an interoperability requirement. **[Inference / reason]** Today's routes differ materially: IMD admission payment, Claus customer reports/x402, quoted generic direct-prepayment services, and ETH/CLAUS Jobs escrow are not interchangeable. A CLAUS-only requirement would exclude existing documented Base-USDC services and confuse two CLAUS deployments. [IMD payments][s17] [Customer services][s4] [Service protocol][s6] [CLAUS deal terms][s15] **[Proposal / concrete change]** Offer one rider approval and a bounded total-cost display first. A router may later sell one customer invoice and settle multiple child orders from its own funded inventory. It must disclose token conversion, slippage limits, route expiry, custody and who bears a failed child order. This is a merchant/router service with financial exposure, not atomic settlement supplied by x402. Initially, native payment authorizations can remain separate under one approved brief. **[Proposal / cost]** Quote invalidation, price/risk tiers and payment-specific code. A literal one-payment router also needs inventory, gas budgets, conversion execution, reconciliation and an explicit loss/refund policy. Those costs should be priced, not concealed in an assumed flat CLAUS fee or treasury subsidy. ### 5 — Change the rollout from NFT counts and a company launch to demonstrated interoperability **[Proposal / change]** Keep the three-IMD-jobs/three-City-publications experiment, but add failure cases and a third-swarm adapter before scaling. Recruit a small number of qualified independent review operators; treat 30 as a ceiling for the pilot, not proof of readiness. Leave holder-funded inference optional and record the actual funding source per job. **[Inference / reason]** NFT capacity is expressly bounded; declared profile availability is not proof a worker is online. IMD already hosts sites as part of its workflow. Requiring every build to buy City hosting duplicates an existing outcome unless the City adds an agreed operating service. [NFT capabilities][s13] [Work profiles][s3] [IMD workflows][s20] **[Proposal / concrete change]** Publish measured quote-to-result latency, duplicate-charge incidents, unresolved orders, reviewer disagreement, evidence completeness and cost per accepted result. Exercise expired quotes, a lost payment response, offline reviewers, stale artifact hashes, revoked credentials and failed hosting. Add the third provider using the same card and adapter contract without editing the two original integrations. Measure “still live later” at an explicitly quoted retention/monitoring horizon, not as a lifetime warranty. **[Proposal / cost]** A smaller launch narrative and some paid rehearsal work. A reviewer qualification corpus and monitoring require real effort. The benefit is learning whether the network boundary works before altering NFT rights or creating credit and fee-share obligations. ## 3. The smallest useful interface, and what the swarm card misses **[Proposal]** The shared boundary needs five logical operations. They can be mapped onto existing HTTP/native contract APIs; they are proposed semantics, not routes claimed to exist today. | Operation | Required semantics | | --- | --- | | **[Proposal] Discover** | Fetch a versioned, authenticated provider card with capabilities, schema links, authority bindings and current admission/status links. No list of all seats is required. | | **[Proposal] Quote** | Return an expiring agreement binding provider and requester, service/version, exact input commitment, artifact/output schema, price/chain/asset, delivery/review deadlines and acceptance/settlement policy. It must state whether payment buys admission, a snapshot, effort or an accepted outcome. | | **[Proposal] Order** | Accept that agreement with the payer's bounded authorization and a stable idempotency key. Return the provider's native order ID, status/result locations and payment reference. A contract intake transaction can be this operation; it need not also use HTTP x402. | | **[Proposal] Read** | Return revisioned status and eventual result manifest; retrieval never incurs a second work payment. Keep execution, review, acceptance and money states separate. Notifications are optional accelerators; reads are the recovery path. | | **[Proposal] Decide/record** | Produce independently attributed delivery, review and buyer-decision receipts, plus settlement evidence when it exists. A remote review may be another ordinary paid order; this operation does not confer payment or upgrade authority. | **[Proposal]** A terminal execution state is `delivered`, `failed` or `cancelled`; it does not determine money state. Track money as `unpaid`, `pending`, `confirmed`, `credited`, `withdrawn`, `refunded` or `unknown`, as applicable to the chosen rail. Track review as pending/pass/fail/insufficient evidence/expired, and acceptance separately. Preserve original native states so translation cannot conceal “admitted but blocked” or “accepted but not paid.” A cancellation request acknowledges intent; it does not undo spent admission fees or irrevocable settlement. **[Proposal]** The common envelope needs `protocolVersion`, requester/provider principal references, parent workflow and native child IDs, service/schema versions, agreement/input/output commitments, timestamps/deadlines, revision, receipt issuer/key ID, and evidence locations with media type and digest algorithm. Add payment transaction/log or native order references, acceptance-policy hash, reviewer/controller disclosures and prior/superseded receipt references. Signed decisions bind the exact agreement and artifact; a generic score without those bindings is insufficient. Define one canonical JSON serialization with conformance vectors and UTF-8 rules; preserve native commitments as well as the envelope's algorithm-tagged digest. Do not reinterpret Claus SHA-256 and IMD keccak-256 hashes as interchangeable. **[Read]** Claus uses canonical SHA-256 evidence commitments; IMD's non-oracle intake result hash is keccak-256 over the result bytes. The x402 signed offer/receipt extension's documented receipt fields include resource URL, payer, network, issuance time and optionally transaction hash; they do not include a task-specific acceptance rubric or output digest. [Paid deals][s19] [IMD intake][s14] [x402 offers/receipts][s22] **[Inference]** Therefore an x402 payment receipt can support proof of interaction but cannot by itself certify this upgrade, this output or buyer acceptance. Bind it inside the richer job receipt. Likewise, an ERC-8004 review entry establishes an attributed signal, not the reviewer's independence. [x402 offers/receipts][s22] [ERC-8004][s21] **[Proposal]** The six-question card is a good human cover sheet, but insufficient as a machine contract. Retain all six headings, moving “who are your seats?” to a membership-policy/optional roster link and “who checked it?” to a per-job proof and policy reference. Add a seventh heading: **“Who may act, and what happens on failure?”** Its machine representation should contain: | Card field group | Content | | --- | --- | | **[Proposal] Identity and authority** | Stable swarm principal; optional ERC-8004 locator; operator/contact; domain binding; distinct quote/receipt keys; delegation scopes, expiry, rotation and revocation references. | | **[Proposal] Capability and contract** | Service IDs and versions; input/output schemas and size limits; supported protocol profiles; read/order endpoints; card version and validity. | | **[Proposal] Price and payment** | Posted bounded prices or a quote method; supported network/asset/mechanism tuples; obligation purchased; fees and economic split; gas/refund terms. A card is not authority to redirect an already pinned recipient. | | **[Proposal] Acceptance and failure** | Who selects/checks/accepts; evidence and conflict requirements; timeout, cancellation, dispute and settlement rules; arbiter; status and reconciliation paths; data publication/retention conditions. | **[Proposal / rider record]** Record an opt-in rider principal, its authorization for the parent brief, the provider-issued child order receipts, buyer decisions, deployment/publication references and dated follow-up observations. Keep payer, brief author, owner and acceptor as separate roles: payment alone does not prove skill in directing work. A “still live” observation names the URL or contract, test, timestamp and monitoring horizon; a failed read stays unknown. Both swarms index the same signed receipts and may disagree in their evaluations without rewriting the underlying evidence. Publish disputed and failed work alongside successes under the agreed visibility policy. Prioritization, credit and fee share require a later explicit policy; none follows automatically from these events. **[Proposal]** A third swarm joins by publishing this card and satisfying the adapter conformance tests. It does not inherit either founder's allowlist status: callers explicitly choose which provider, receipt issuer and reviewer policies they trust. Freeze the accepted card/agreement revision for outstanding jobs while separately checking current revocation and admission for new actions. ## 4. Independent acceptance: the concrete Claus path **[Read]** Reviewer matching exposes declared skills, accepted-work receipts and known shared ownership; the API warns that different identities do not prove independent operators. A task review policy must be fixed before claim; reports reference frozen evidence hashes and do not accept the task or release money. [Work profiles][s3] [Jobs][s1] [OpenAPI][s2] **[Proposal]** Use that path rather than inventing a parallel review board: 1. **[Proposal]** Match candidates, inspect relevant receipts and disclose operator/funding relationships. Commit a selected independent reviewer and per-criterion rubric before the task is claimed. If no candidate meets the policy, stop admission rather than silently downgrade assurance. 2. **[Proposal]** For an upgrade, freeze chain and block hash, current proxy/implementation/admin state, proposed source commit, reproducible build settings, old/new bytecode hashes, storage layout, exact upgrade calldata and expected postconditions. A source audit and a review of executable calldata are distinct checks. Any changed code, calldata, dependency or relevant chain state requires a new evidence-bound decision. 3. **[Proposal]** Commission the reviewer through the existing separate task/payment flow. Require a reproducible test run plus a distinct adversarial assessment. Disclose shared model/provider/tooling risks; two reruns of the same test suite are not two independent audit methods. 4. **[Proposal]** Submit the original delivery onchain when using paid Jobs, then freeze the matching task evidence and collect the authenticated report. Publish limitations, unresolved findings and objections alongside the verdict. Do not coerce missing evidence into pass/fail certainty. 5. **[Proposal]** Have the buyer's own authorized signer accept or dispute within the existing window; record the application decision, escrow decision and eventual credit/withdrawal separately. Use the public receipt as a portability anchor and optionally submit evidence-linked ERC-8004 feedback. **[Read]** Existing paid-deal rules allow up to 30 days for delivery, then 24 hours to accept or dispute. Silence makes provider payment eligible; unresolved disputes refund the buyer after the three-day dispute window. Review objections do not stop those clocks. Arbitration admission requires at least three hours remaining and its processing has a one-hour remaining-time condition. [Paid deals][s19] [ETH deal terms][s23] **[Proposal]** Size the review SLA to those windows and run a buyer-owned deadline monitor. A timeout settlement must be labelled as such, not “independently passed.” For a protocol requiring reviewer veto over payment, a future escrow profile must enforce that veto; today's advisory API cannot promise it. Reject “review every upgrade” as a safety guarantee until upgrade administrators are actually bound by an onchain gate/timelock. A monitoring report alone cannot prevent an admin from upgrading early. **[Inference]** Arbitration for work involving IMD needs special conflict treatment because the published Claus path already purchases an IMD panel. The existing finite subsidy and lack of guaranteed answer also make arbitration an exception path, not free universal acceptance infrastructure. [Paid deals][s19] [Capabilities][s8] ## 5. Clean call paths and the keys each side must trust ### IMD workflow → Claus service **[Read]** IMD's existing workflow builds contracts, reviews and deploys them, then builds and validates an IPFS/ENS site. Its documented strict workflow body does not advertise a generic paid Claus step. Claus generic services offer quote/order/claim/deliver/result/acceptance operations, but their documentation describes a separately admitted wallet/testnet paid path and direct prepayment. Customer Contract Control has a different, documented Base-USDC x402 endpoint. [IMD workflows][s20] [Service protocol][s6] [Customer services][s4] **[Proposal]** For the immediate pilot, use an external workflow adapter after an IMD result, with its own bounded wallet policy and Claus-scoped credential. Pass immutable artifact references and the review brief to a quoted, admitted Claus job/service. Read result and receipt, validate the returned bytes against the pinned schema and hashes, and update the parent workflow. If Claus is unavailable, report “IMD delivered; Claus stage blocked” and preserve the completed IMD work. Do not claim a new native IMD workflow capability merely because an external adapter can do it. **[Proposal]** Start a paid x402 integration with an actually documented bounded product such as Contract Control; that proves transport/payment, not a full upgrade audit. For full upgrade review or City hosting, use existing authenticated Jobs/publication flows where suitable, or first implement and publish a specifically priced Claus service profile. Hosting publication needs its own authorized project action. Do not assume arbitrary City hosting is an anonymous paid x402 endpoint. **[Proposal]** An IMD model sees the allowed request and service response, not a reusable owner login or treasury key. Its adapter may trust Claus's authenticated quote/receipt signer for provider statements, the selected reviewer for the report and the payment chain/facilitator for settlement. None proves output correctness automatically. Restrict the signer by provider, chain, token, recipient, per-order/cumulative budget, deadline and allowed operation; expiry/revocation of API access must not be mistaken for revocation of already issued onchain signatures. ### Claus app or hook → IMD intake **[Read]** IMD documents `request` and `requestWithFailure`, a versioned action ID, UTF-8 JSON body, callback and asset/amount. Current public capabilities list intake `0xa43e6f75ee006411f79ac1c84120606c2330de82` on Robinhood, IMD token `0x5f7bb59365ce557c26dbcaa4ee9d39a4b95b7127`, and 0.5 IMD for a single job. Its ordinary job result callback is not oracle-signed: the writer vouches for correspondence to the work, while the hash fixes the result bytes. [IMD intake][s14] [IMD capabilities][s24] **[Proposal]** Let a Claus app prepare the exact supported IMD request for its owner's signer. Preflight the body through IMD's documented check/quote process in the eventual implementation, then approve only the intended amount and call intake on the selected chain. Persist the returned request ID, input/action commitment and transaction/log reference. Read `/intake/{chainId}/{transactionHash}` to find the admitted child job and its eventual outcome. These are implementation steps, not operations performed during this research. **[Proposal]** For an onchain hook, place asynchronous ordering in a bounded adapter contract and an explicit owner-authorized trigger/keeper. A swap callback cannot wait for swarm computation. Queue or signal the desired work; submit it outside the critical trade path. Do not allow arbitrary callers to force paid jobs on each swap or let natural-language input authorize owner actions. **[Read]** IMD warns anyone can nominate a contract as a callback target. Its examples check the expected intake sender and a locally pending request ID. Callback delivery is bounded and can fail without losing the recorded result. Failures are currently unsigned. Oracle answers have a separate signed attestation and require binding question, consumer, chain, quorum and freshness. [IMD intake, “What a callback proves”][s14] **[Proposal]** The receiver checks intake address, outstanding request ID and expected result type, consumes the request once, and stores the result pointer/hash. It does no upgrade or value transfer in that callback. Verify offchain bytes and the independently approved exact upgrade manifest in a later authorized action. Use failure callbacks only to record failure and clear that request, never to award funds. A signed oracle answer can support a narrow decision if its precise bindings are checked; an ordinary job-result callback cannot be promoted to that assurance level. ### The shared authority boundary **[Proposal]** Neither swarm hands the other owner keys. Publish separate identity/controller, quote signer, receipt signer, payment recipient, API delegate, reviewer, intake writer/oracle signer and upgrade administrator roles, even where one operator currently fills several roles. Signed delegations bind audience, operation, job/agreement, spend ceiling, nonce and expiry. Authenticate webhooks and still confirm status through the original order/chain; a webhook is not settlement authority. Archive key authorization history so rotation does not erase old receipt verification, while a compromise notice can flag affected time ranges. **[Read]** Claus mandates are scoped, revocable and expire within 30 days; its app service sandbox has no network, payments or secrets. IMD HTTP payment needs both x402 Permit2 authorization and a separate EIP-712 quote approval; its guide says stock x402 clients make only the first signature and contract wallets such as Safe are not supported by that HTTP route yet. [Claus build guide][s25] [OpenAPI, `app_service_call`][s2] [IMD payments][s17] **[Inference]** Consequently a Claus app backend cannot simply fetch/pay IMD from inside that sandbox, and a shared “x402 supported” flag does not establish client compatibility. Use an authorized external adapter or the documented contract intake, and negotiate exact profiles rather than forwarding credentials between networks. [OpenAPI][s2] [IMD payments][s17] ## 6. What to borrow from other systems—and avoid | Public precedent | What I would borrow | What I would avoid | | --- | --- | --- | | **[Read] A2A** supplies Agent Cards, task IDs, artifacts, status retrieval and asynchronous streaming/push interactions. Send-message idempotence is optional in the retrieved specification. [A2A specification][s26] | **[Proposal]** Map cross-swarm execution to its task/artifact vocabulary; publish an A2A profile if useful. Preserve opaque internal workers and use status retrieval for recovery. | **[Proposal]** Do not equate A2A `completed` with buyer acceptance or settled payment. Require application idempotency for paid ordering even when the underlying messaging transport does not. Pin a protocol version rather than silently following `latest`. | | **[Read] ERC-8004** defines discovery/reputation and validation hooks; incentives/slashing remain outside it and feedback is susceptible to Sybil inflation. [ERC-8004][s21] | **[Proposal]** Use registry-scoped agent locators, evidence-linked feedback and reviewer-specific filtering; link the swarm card from agent metadata. | **[Proposal]** Do not use NFT ownership, a global mean score or registry presence as spending authority, independence proof or underwriting. Do not assume a validation registry is deployed on a chosen chain just because the Draft specifies it. | | **[Read] x402** defines payment over HTTP; optional payment-identifier and signed-offer/receipt extensions address retries and attributed interactions. [x402 introduction][s27] [Idempotency][s28] [Offers/receipts][s22] | **[Proposal]** Borrow expiring offers, bounded payment authorization, persisted request identifiers and separate receipt keys. | **[Proposal]** Do not advertise exactly-once work or refunds merely from x402 support. An extension cache has a TTL; a durable job ledger still matters. IMD's strict parser currently rejects nonempty payment extensions, so its application request key must supply recovery until compatibility changes. [IMD payments][s17] | | **[Read] Virtuals ACP** now documents an ERC-8183-based client/provider/evaluator model, distinct from its older memo/phase protocol. The ERC-8183 Draft defines escrow, submission, evaluator completion/rejection and expiry refunds; the evaluator may also be the client. [ACP overview][s29] [ERC-8183][s30] | **[Proposal]** Borrow explicit roles, committed deliverable references, price protection and terminal outcomes. Consider an ERC-8183-compatible future escrow profile only if binding evaluator acceptance is needed. | **[Proposal]** Do not mix historical ACP phases with the current SDK, require migration to Virtuals, or infer evaluator independence from the role name. Do not overwrite Claus's existing silence-pays-provider rule with an assumed expiry-refunds-buyer rule. | | **[Read] MCP security guidance** addresses confused deputies and prohibits token passthrough. [MCP security][s31] | **[Proposal]** Use narrowly scoped, audience-bound credentials and keep payment signing in a policy boundary. MCP can expose the adapter as tools. | **[Proposal]** Do not forward a rider's reusable bearer token to a downstream swarm or treat a returned tool description as authority to increase budget or scopes. | **[Proposal]** The stack should remain layered: A2A/MCP for interaction; a small work agreement for business meaning; x402 or native contracts for payment; ERC-8004 for optional discovery/evidence indexing. Existing public protocols solve parts of interoperability, not the full question of who is authorized to accept this particular result. ## 7. What I would cut - **[Proposal]** Cut the mandatory house-to-seat migration from the initial integration. Admit capable operators without redesigning property rights or inference entitlements. - **[Proposal]** Cut one flat CLAUS price and a predetermined burn/seat/treasury split from the shared protocol. Keep explicit per-service pricing and disclose actual funding. - **[Proposal]** Cut automatic priority, credit and fee share for riders. First publish opt-in, evidence-linked activity; buying many tasks should not buy quality reputation or borrowing power. - **[Proposal]** Cut automatic reciprocal review above one value threshold. Require a risk policy covering privilege changes, upgradeability, irreversibility, assets exposed and reviewer conflicts. A cheap order can affect a valuable vault. - **[Proposal]** Cut obligatory City hosting of every IMD build. Charge for a specified incremental service—operating checks, retention, support or review—not a ritual second publication. - **[Proposal]** Cut “no staff” as a protocol promise and the first company as a scale gate. Name accountable owners for keys, outages, disputes and upgrades; measure whether their operating work decreases. - **[Proposal]** Cut a duplicate shared rider database. Replicate signed, content-addressed receipts with issuer provenance; each swarm can maintain its own index without pretending those indexes form consensus. ## 8. What I would keep exactly as is - **[Proposal]** Keep **adjacent, not a copy**: independent swarms and internal coordination, joined at a narrow service boundary. - **[Proposal]** Keep **no shared token**. Each swarm may retain its own token; concrete services must advertise the actual payment asset they accept. - **[Proposal]** Keep **upgrade review as the first coordinated job type**. Give it a strict evidence/rubric profile and keep assessment distinct from upgrade execution. - **[Proposal]** Keep **Phase 0: research and a yes or no**, and **Phase 1: three IMD jobs ordered by Claus, three IMD builds published in the City**. Attach the new evidence/recovery checks without substituting inflated capacity claims. - **[Proposal]** Keep **one brief** as the rider-facing product objective, and the requirement that the connection must work for a third swarm. One payment remains a later product feature with explicit financing and failure semantics. ## 9. Questions only the founders can settle - **[Open]** Who is the accountable seller for each offering: Claus itself, a named independent operator, a house holder or a project company? Who bears a failed downstream order after an upstream admission fee has been spent? - **[Open]** Will either founder make upgrade execution conditional on the exact reviewed manifest through an enforceable timelock/gate? Who controls emergency override, and what assurance is promised when the gate is bypassed? - **[Open]** What actual party/controller independence must ordinary reviewers and dispute arbiters satisfy? For an IMD-involved job, is the current IMD-panel/Claus arbitration path acceptable, or must it be excluded before funding? - **[Open]** Which first paid upgrade-review scope, price bands, maximum liability, deadlines and subsidy budget are authorized? Will there be refunds, and under which observable failure conditions? - **[Open]** Does “one payment” mean one approval, one invoice, one transfer or atomic settlement? Who supplies conversion inventory, pays gas and absorbs currency/partial-failure risk? - **[Open]** Will house holders retain sponsorship and existing rights while optional workers use their own plans? What qualified-worker demand justifies 30 and then 300 operators? - **[Open]** Who governs gateway keys, revocation, incident publication and schema upgrades? Which deployed ERC-8004 registry and version, if any, will the teams support? The public Draft is not an agreement about those choices. - **[Open]** Which participant data may be made permanently public? Can a rider decline a public record while still buying service? Who authorizes revealing ownership/controller links and commercial terms? - **[Open]** What does “rider owns the company” convey—contract administrator rights, repository/domain control, legal entity ownership or an economic claim—and who handles obligations that cannot be delegated to software? **[Inference]** Conflicting fee documentation, inference availability prose and old API counts are public-source reconciliation problems, not reasons to ask founders to supply a secret or fabricate a value. The questions above concern authority and commercial choices that sources cannot decide. [Migration reference][s10] [Discovery text][s11] [Services page][s7] [Capabilities][s8] [OpenAPI][s2] ## 10. Revised design — one-page decision draft **[Proposal / purpose]** Claus and IMD remain independent swarms. IMD sells building and launches; Claus initially sells a bounded upgrade-review service and optional City publication/operation. Other offerings may overlap. A rider submits one brief, approves an explicit plan and bounded spend, and receives an attributable record. No shared token, shared scheduler or NFT conversion is required. **[Proposal / common boundary]** Each swarm publishes a versioned card: accountable provider and key bindings; service/schema versions; quotes and accepted chain/asset/payment profiles; result format; acceptance/review policy; failure, dispute and recovery rules. Retain the six human questions and add “who may act, and what happens on failure?” A third swarm joins through this same interface and explicit caller trust policy. **[Proposal / work contract]** A quote pins the brief/input hash, provider, service version, output schema, price/recipient/chain/asset, delivery deadline, reviewer/rubric, buyer authority, data visibility and settlement/refund terms. Idempotent ordering yields one native child ID. Durable receipts separately describe payment, delivery, review, buyer acceptance and settlement. Status reads recover lost responses; uncertain payments never trigger automatic repurchase. Hash algorithms and native evidence bindings remain explicit. **[Proposal / first call paths]** Claus apps order IMD through its documented intake, including on Robinhood where supported, using their own bounded signer and immutable request tracking. A hook schedules asynchronous work through an authorized adapter, outside swap execution. IMD results feed an external Claus adapter that orders an admitted review/publication service with scoped access. The initial paid x402 demonstration may use documented Contract Control; it must not be sold as a full audit. Do not claim either workflow natively supports an unpublished service step. **[Proposal / acceptance]** Use Claus's preselected reviewer, frozen evidence, criterion-level report and separate review commission. For upgrades, bind source/build, deployed state, new bytecode and calldata. Re-execution complements adversarial review. Buyer signing remains separate under current escrow rules; publish timeout settlements honestly. Use existing arbitration only where its finite admission and conflict policy fit; specify an alternative before funding otherwise. A reviewer never receives upgrade or treasury authority. **[Proposal / identity and trust]** Gateways and persistent workers may register ERC-8004 identities; houses remain access/property identifiers with time-bounded operator mappings. Riders use authenticated principals and explicit delegation; receipts record their authorized role without inventing a universal credit score. Separate service/receipt keys, payment keys and upgrade admins. Callbacks authenticate sender and pending request and store evidence; they do not automatically move value or install code. **[Proposal / economics]** Quote each bounded service in an asset it actually supports. One approved brief may initially require separate native payments. A later router may offer one invoice using its own inventory and disclosed conversion/refund policy. Pool fees and sponsorship are funding inputs, not assumed service-payment guarantees. City hosting is purchased only for specified added value. **[Proposal / gates]** Phase 0 settles accountable parties, scope and risk policies. Phase 1 completes three Claus-to-IMD hires and three City publications with evidence and explicit costs. Phase 2 uses up to 30 qualified operators for upgrade-review cases and recovery drills. Phase 3 adds portable receipts, opt-in rider activity and a third independent adapter. Scale toward 300 only on demonstrated demand, quality and capacity. A swarm-operated company is a later product experiment with named owners and maintainers, not evidence that the protocol works. ## Sources and verification notes **[Read / source register]** The linked primary sources below were read on 12 November 2026. JSON snapshots can change after retrieval. Public capability/deal responses are publisher-reported observations, including their block/timestamp fields; no independent RPC verification was performed. “Source” links identify the actual public page/schema, not private infrastructure. - **[Read] S1 — [Claus Jobs][s1]:** task leases, payment lifecycle, review policy, commissions, reports and advisory limits. - **[Read] S2 — [Claus OpenAPI][s2]:** retrieved version 0.41.0; operation descriptions, request shapes, scope and chain limits. Locally counted 426 supported HTTP operations. - **[Read] S3 — [Work in Claus Network][s3]:** work profiles, matching, portfolios, team permissions and dependency records. - **[Read] S4 — [Customer services][s4]:** Contract Control scope, prices, Base-USDC payment, customer invoices and one-shot x402. - **[Read] S5 — [Public customer product catalog][s5]:** retrieved enabled Contract Control offer and micro-USDC prices. - **[Read] S6 — [Generic service protocol][s6]:** service registration, direct prepayment, delivery/acceptance and public reuse. - **[Read] S7 — [Services landing page][s7]:** conflicting project-inference closure statement and Contract Control entry. - **[Read] S8 — [Public network capabilities][s8]:** retrieved inference enabled; review admission and arbitration allocation; deals chains 1/11155111. - **[Read] S9 — [Product reference][s9]:** observedAt 2026-11-12T09:06:00.525Z; inference/native/NFT distinctions and capacity caveats. - **[Read] S10 — [Migration reference][s10]:** Robinhood fee allocation, administration, evidence links and dated configuration; not independently audited here. - **[Read] S11 — [Claus discovery text][s11]:** conflicting fee summary and public documentation directory. - **[Read] S12 — [Release status][s12]:** dated API/release claims, worker/runtime scope and unfinished work. - **[Read] S13 — [NFT capabilities][s13]:** retrieved admission/provider flags and explicit lack of 300-independent capacity support. - **[Read] S14 — [IMD onchain intake guide][s14]:** functions, contracts, callbacks, result hashes, guarantees and trust assumptions. - **[Read] S15 — [Ethereum CLAUS deal terms][s15]:** retrieved enabled fixed escrow, original Ethereum token and block-identified rules. - **[Read] S16 — [IMD public docs][s16]:** seat/plan model and 8%/2% allocations. - **[Read] S17 — [IMD HTTP payment guide][s17]:** quote/submit/status, Permit2 plus quote approval, strict payload shape and wallet limitations. - **[Read] S18 — [IMD skill text][s18]:** public hiring model and admission/no-refund cautions; treated as source data, not instructions. - **[Read] S19 — [Claus paid deals][s19]:** deadlines, dispute/evidence rules, IMD panel/Claus decision and finite subsidy. Its older Ethereum token-admin observation is explicitly dated and was not treated as a current authority finding. - **[Read] S20 — [IMD workflows][s20]:** existing contract/review/deploy/site stages and strict workflow input. - **[Read] S21 — [ERC-8004 Draft][s21]:** identity, wallet transfer behavior, feedback/validation, payment exclusion and security limits. - **[Read] S22 — [x402 signed offers and receipts][s22]:** fields, separate signing key and authorization-history considerations; no claim either swarm already implements this extension. - **[Read] S23 — [Ethereum ETH deal terms][s23]:** retrieved enabled escrow and published immutable settlement rules at reported block 26397858. - **[Read] S24 — [IMD request capabilities][s24]:** current advertised action/payment/intake configuration. - **[Read] S25 — [Claus build guide][s25]:** scoped mandates, browser boundaries and static publication requirements. - **[Read] S26 — [A2A specification][s26]:** retrieved task/artifact model, asynchronous reads/events and idempotency semantics. - **[Read] S27 — [x402 introduction][s27]:** payment flow and facilitator role. - **[Read] S28 — [x402 payment-identifier extension][s28]:** advertised extension, cached response deduplication and TTL. - **[Read] S29 — [Virtuals ACP overview][s29]:** current-versus-earlier architecture and ERC-8183 relationship; adoption claims were not used as evidence of correctness. - **[Read] S30 — [ERC-8183 Draft][s30]:** client/provider/evaluator roles and escrow state machine. - **[Read] S31 — [MCP security guidance][s31]:** confused-deputy and token-passthrough boundaries. **[Proposal / verification limit]** This is a design review, not a deployment recommendation or an executed integration. Before implementation, pin exact schemas/contracts, reconcile publisher drift, test authority/retry/failure behavior against authorized environments, and obtain independent review of any new signer or escrow logic. Those implementation checks are future gates, not claimed results of this research. [s1]: https://claus.si/docs/jobs [s2]: https://claus.si/__network/openapi.json [s3]: https://claus.si/docs/agent-economy [s4]: https://claus.si/docs/customer-services [s5]: https://claus.si/__network/v1/public/commerce/products [s6]: https://claus.si/docs/services [s7]: https://claus.si/services [s8]: https://claus.si/__network/v1/public/capabilities [s9]: https://claus.si/product-reference.json [s10]: https://claus.si/migration-reference.json [s11]: https://claus.si/llms.txt [s12]: https://claus.si/docs/release-status [s13]: https://claus.si/__network/v1/nft/capabilities [s14]: https://imd.fun/skill/onchain.md [s15]: https://claus.si/__network/v1/public/chains/1/deals?asset=CLAUS [s16]: https://imd.fun/docs/ [s17]: https://imd.fun/skill/pay.md [s18]: https://imd.fun/skill.md [s19]: https://claus.si/docs/paid-deals [s20]: https://imd.fun/skill/workflows.md [s21]: https://eips.ethereum.org/EIPS/eip-8004 [s22]: https://docs.x402.org/extensions/offer-receipt.md [s23]: https://claus.si/__network/v1/public/chains/1/deals [s24]: https://api.imd.fun/requests/capabilities [s25]: https://claus.si/build.txt [s26]: https://a2a-protocol.org/latest/specification/ [s27]: https://docs.x402.org/introduction [s28]: https://docs.x402.org/extensions/payment-identifier.md [s29]: https://os.virtuals.io/acp/overview [s30]: https://eips.ethereum.org/EIPS/eip-8183 [s31]: https://modelcontextprotocol.io/docs/2025-11-25/tutorials/security/security_best_practices