# Claus beside IMD: federate orders, not seats Researcher 3 · Protocol and interface lens · Public-source reading, 12 November 2026 **[Proposal] Verdict:** proceed with a small procurement federation. Replace the proposed seat conversion and bilateral “bridge” with versioned orders, explicit settlement terms, and portable evidence. Claus should remain a distinct operating and review environment. A third swarm should implement one boundary interface, without adopting either network’s coordinator, NFT economics, or model subscription policy. **Reading key:** **[Read]** means a claim explicitly made by the linked public source, not independently certified behavior. **[Inference]** means my interpretation of that evidence. **[Proposal]** means my recommendation, including suggested thresholds and interface fields. A label governs its whole paragraph, list, or table unless another label intervenes. **[Open]** identifies a founder decision or unresolved factual question. The supplied v0.1 is the object of criticism, not independent evidence. **[Read / research limits]** I read public pages, public machine-readable schemas and capability responses, IMD’s published skill references, and standards documentation. Some pages unavailable through the research browser were readable with ordinary public HTTPS GETs. I did not authenticate, invoke mutation endpoints, submit quotes, purchase services, inspect private systems, or read other researchers’ views. Consequently this report establishes documented interfaces and observed declarations, not end-to-end operation, actual operator independence, deployed contract safety, or paid customer adoption. All linked sources were consulted on the report date; mutable documents were not archived or pinned to repository commits. ## Correct the baseline first | Status | Correction to v0.1 | Design consequence | | --- | --- | --- | | **[Read]** | Claus already documents independent task reviews: a reviewer and criteria committed before claim, frozen input/output and evidence hashes, immutable criterion reports, objections, and a separate paid review commission. Review remains advisory; requester acceptance and payment are separate. [C1] | **[Inference]** “No independent acceptance” conflates independent assessment with authority to accept or release money. Extend the existing review policy rather than inventing another task-review system. | | **[Read]** | The public OpenAPI includes `reviewer_candidates`, which matches declared skills, shows accepted-work receipts, flags known shared ownership, and warns that separate identities do not prove independent operators. [C2] | **[Inference]** Matching infrastructure already exists; the unresolved problem is independence policy and verification of competence. | | **[Read]** | Claus documents disputed-deal arbitration using one IMD oracle request with five panelists and four agreeing, zero numerical tolerance, followed by Claus approval of the exact split. A completed case yields an expiring resolution signature; a party still submits settlement. [C3] | **[Inference]** Cross-network review is already a documented integration, not solely a phase-3 invention. It also creates a dependency that complicates claims of independent cross-review. | | **[Read]** | Jobs’ paid agreements use Ethereum wallets; the public deal schema offers chains 1 and 11155111, not 4663. CLAUS selection identifies a specific escrow and token. Separately, Claus’s home page identifies its current official coin on Robinhood Chain. [C1, C2, C4] | **[Inference]** Do not treat an Ethereum CLAUS escrow and the current Robinhood CLAUS as interchangeable settlement assets. A token symbol cannot identify payment. | | **[Read]** | The NFT guide describes 300 houses and sponsored allowance up to $5/NFT/day, with automatic top-up off. It expressly says a key does not start an always-running agent. Transfer does not move private projects, account history, or keys. [C5] | **[Inference]** Houses are not presently proven working seats, but neither are all Claus agents reducible to houses. Turning property ownership into worker identity would alter existing transfer semantics. | | **[Read]** | Public capabilities declare workers, durable runtimes, workflows, matching, and infrastructure. These are bounded execution profiles, not evidence of a decentralized seat scheduler. [C6] | **[Inference]** The blanket gap “no coordinator” is too strong. Ask which coordination guarantees are missing instead of presuming Claus starts with hosting alone. | | **[Read]** | The HTML API reference says 395 operations. Counting method objects bearing `operationId` in the public OpenAPI returned 426 during this reading. [C2, C7] | **[Proposal]** Publish schema version and digest; never use operation count as a compatibility promise. | | **[Read]** | The inference guide and build guide say new project inference submissions closed on 21 October 2026. The capability response instead reported `inference.enabled: true`. NFT inference is documented separately. [C5, C6, C8, C9] | **[Open]** This is a public-source conflict, not proof of reopening. Treat project inference admission as unresolved and exclude it from the pilot’s dependencies. | | **[Read]** | Customer services document a Contract Control x402 endpoint with Base USDC, alongside saved direct-payment invoices. The generic service guide describes a separately admitted Sepolia payment path. [C10, C11] | **[Inference]** “Claus x402 services in USDG” is not a safe blanket description of the current interface. Discover each product’s chain, asset, method, and admission individually. | **[Read]** IMD’s public hiring guide states 0.5 IMD/request, owner-run machines and model plans, review by other agents, and admission rather than a guaranteed result. Its onchain guide documents Ethereum and Robinhood intake and asynchronous callbacks, with no refunds. [I1, I2] **[Inference]** Those parts of v0.1 have a public basis, but there is no common delivery guarantee to inherit merely by connecting the two networks. **[Read]** IMD’s docs state 2,000 seats, an 8% launch allocation to connected seats and 2% to wallets with accepted work. [I3] **[Proposal]** Preserve these as IMD’s internal economics, not federation requirements. I did not independently establish IMD’s deployed ERC-8004 registration mapping; the federation must publish actual registry identifiers before claiming compatibility. **[Open]** I did not obtain Matt Huang’s original dated post from a primary source. Its exact wording and date remain unverified here. **[Proposal]** Treat the rider thesis supplied in the brief as a product hypothesis; no protocol or credit decision below depends on the post being authentic or proving that companies require no human staff. ## Five changes, ranked ### 1. Standardize the order and its failure semantics before standardizing the swarm **[Proposal] Change:** require a small versioned procurement profile with discovery, quote, submit, read status/result, and record review/decision. Implement adapters to existing APIs. A common coordinator, seat NFT, token, or global registry is unnecessary for the first order. **[Inference] Reason:** v0.1 describes who does work and which coin pays, but leaves the hardest interoperability questions implicit: what payment purchases, what constitutes delivery, which actor may accept it, and what happens after a lost response. Existing Claus and IMD documentation already distinguish these events. [C1, C11, I2] **[Proposal] Smallest common messages and obligations:** | Message | Required binding | Required behavior | | --- | --- | --- | | Service descriptor | Provider ID; service/version; input/output schemas; authenticated endpoint; capability expiry | Discover a capability without enumerating every seat. | | Quote | Buyer; provider; service/version; input digest; terms digest; price ceiling; chain and token address; payee; expiry; delivery deadline; acceptance/dispute policy | Say whether payment buys admission, prepaid execution, or escrow-backed delivery. | | Order | Buyer-generated idempotency key; quote digest; payment reference; provider order ID; parent order reference if delegated | Same key plus same bytes returns the same order; different bytes conflict. | | Delivery | Order reference; artifact URI and digest; media type; provenance; delivered-at time; retention commitment | Make the exact result retrievable and verifiable. A URL alone is insufficient. | | Review and decision | Exact order, delivery and rubric digests; assessor; verdict; evidence; conflicts; limits; expiry; decision authority | Record assessment separately from customer acceptance and financial settlement. | **[Proposal]** Keep execution, review, and finance as separate status dimensions. Execution can be queued, running, delivered, failed, or cancelled; review can be absent, pending, pass, fail, or inconclusive; finance can be unconfirmed, paid, credited, withdrawn, refunded, or unknown. Terminal execution failure must not erase payment uncertainty. Derive a user-facing next action from those dimensions, not a single misleading `completed` flag. **[Proposal]** Polling is mandatory; authenticated webhooks are optional hints. A lost webhook cannot lose work. Retain original order/payment identifiers, expose machine-readable terminal errors, and prohibit automatic replacement purchases while settlement is unknown. Cancellation stops future execution only to the extent the quote promises; it cannot reverse an already settled payment or mined transaction. **[Proposal] Cost:** both teams maintain mapping adapters, durable deduplication records, schema fixtures, and retention policy. Some current endpoints will need richer receipts. This adds boundary engineering but avoids requiring either team to rebuild its internal scheduler. ### 2. Extend Claus’s existing review machinery; do not relabel it automatic independent acceptance **[Read]** Claus already binds reports to frozen evidence and selected criteria; a failed or inconclusive paid review is still compensable work. Review objections do not stop escrow clocks. [C1] Its paid-deal rules give a buyer 24 hours after delivery to accept or dispute, a three-day dispute window, and a buyer refund if arbitration does not resolve in time. [C3] **[Proposal] Change:** use `reviewer_candidates` for discovery and existing task review/commission operations for evidence and payment. Add a versioned acceptance policy that names three distinct roles: assessor, customer acceptance authority, and settlement authority. For the pilot, independent review advises the requester; it does not acquire upgrade or withdrawal authority. Any later automatic release requires separately reviewed settlement logic and opt-in terms. **[Proposal] Upgrade-review procedure:** 1. Before claim, freeze scope, acceptance rubric, assessor eligibility, fee, deadlines, conflict policy, and appeal route. Identify the target chain, contract, current implementation, proposed implementation and code hash, exact calldata, source commit, storage-layout evidence, and relevant state block/hash. 2. Select reviewers using accepted-work evidence and known-controller exclusions. Record employer/operator/funding disclosures and dependency on IMD. Obtain a different operator and preferably different model/tooling for the second assessment. Do not call a second run by the same controller independent. 3. Run a reproducible build and targeted simulation/tests from the frozen bundle; record commands, tool versions, assumptions, and findings. Separately assess economic effects and authority. A second run can reproduce a shared blind spot. 4. Submit criterion-level pass/fail/insufficient-evidence findings through the existing immutable report path. Pay for review effort, not a favorable verdict. Bind fixes to a new artifact and a new review of that artifact; never carry approval forward by project name. 5. The requester records acceptance, and the authorized owner decides execution. Immediately before executing an upgrade, verify code/calldata/state assumptions and expiry again. Material drift requires new review. Keep timelocks and recovery outside reviewers’ keys. 6. For financial disputes, monitor the original escrow deadline and submit the actual dispute transaction when appropriate. An advisory objection is not a dispute. If review cannot finish in time, expose the documented fallback outcome and its provider risk. **[Read]** Claus’s coin-review page already requires exact code/calldata, current state and authority checks, compatible storage, and an explicit decision before execution; it also acknowledges root-administrator power over protected execution-service rules. [C12] **[Inference]** A report cannot convert an administrator-controlled execution service into an immutable safety boundary. **[Proposal]** Replace “above a set value gets another swarm’s review” with a risk policy: authority-changing, custody-changing, or issuance-changing upgrades always require the stronger review class, even when the purchase price is small. Other jobs use buyer-selected assurance tiers. Record dependency provenance: a Claus opinion derived from an IMD panel cannot count as an independent second opinion on that same panel’s work. **[Proposal] Cost:** more reviewer fees, slower upgrades, reproducible evidence storage, conflict assessment, deadline monitoring, and inconclusive outcomes. Independence is partly governance and disclosure; no NFT or matching API proves it. An automatic settlement revision would add contract review and migration costs, so defer it. ### 3. Separate service identity, worker identity, property, and rider reputation **[Read]** ERC-8004 is a draft with identity, reputation, and validation registries; payments are outside its scope. Identity is an ERC-721 registration pointing to a registration file, globally qualified by chain/registry and agent ID. Transfers affect ownership; the registered payment wallet is cleared on transfer. Feedback is open to clients, with Sybil risks explicitly acknowledged. Validation incentives remain outside the standard. [E1] **[Proposal] Change:** use ERC-8004 as an optional shared agent/service identity anchor, not as a mandatory conversion of all houses or the definition of a human rider. Register the accountable service boundary first; add workers when independently attributable work justifies it. **[Proposal] Register exactly this:** | Subject | Identity and registration contents | Attribution rule | | --- | --- | --- | | Claus procurement/review gateway | Agent registration with official service endpoints, protocol versions, registration references, payment-wallet binding, supported trust profiles, and link to signed swarm descriptor | Makes the gateway accountable; does not certify all workers or administrator behavior. | | Participating Claus worker agent | Separate registration owned by its accountable controller; endpoints or delegation pointer; skills; signed membership and validity interval | Receipts name the agent ID, controller/delegation epoch, and actual role at work time. | | House | Existing property ID linked optionally to worker registrations | A house is a membership/property claim. It does not carry the seller’s human competence or private account into the buyer. | | Rider/customer | Wallet or organization identifier, delegated signing keys and epochs, optional privacy-preserving public alias | Do not mint a transferable “human competence” NFT. A rider’s software gateway may separately register as an agent. | **[Proposal]** Register metadata and public endpoints, never model-provider credentials, private prompts, wallet secrets, or personal customer files. An ERC-8004 identity owner, payment wallet, API credential and upgrade administrator must remain distinct concepts. Publish explicit authorization linking the service’s signing keys and settlement payees; verify those links when paying rather than trusting a logo or URI. **[Proposal]** The rider record is a portable set of signed receipts, not two independently maintained scores. Include buyer, payer, beneficiary, parent/child order IDs, terms/output digests, assessor and acceptance authority, failure/cancellation outcomes, payment evidence, and later health observations with measurement time and method. Preserve distinct roles: paying for a launch is not proof of owning its contracts or doing its work. **[Proposal]** Each swarm may index the same receipts and compute its own ranking. Customer feedback can be exported to ERC-8004 where appropriate; do not disguise customer satisfaction as independent validation. Use a validation adapter only for explicitly specified validator policies, with report linkage and update semantics. Publish no arbitrary numeric score that implies all assurance types are interchangeable. **[Proposal]** Begin rider records with opt-in visibility and a published measurement rubric. Delay priority, credit, and fee shares: purchase volume can be manufactured, buyers can collude with sellers, and a transferable wallet is not a stable human identity. Credit needs an explicit underwriter, loss budget, and recourse policy; it is not an automatic consequence of an onchain history. **[Proposal] Cost:** identity mapping and key-rotation machinery, attribution at ownership changes, indexers, optional registry gas, privacy choices, and compatibility work as a draft evolves. It sacrifices the simplicity of “300 houses = 300 workers,” but preserves the distinction the existing house-transfer rules already make. ### 4. Make payment adapters explicit and keep signing authority local **[Read]** x402 defines HTTP payment negotiation; EVM transfer support includes EIP-3009 and Permit2, with token/network-specific requirements. [X1, X2] IMD’s HTTP flow additionally requires a quote-approval signature from the payment signer; its guide warns that a stock x402 client does not produce that second signature. [I4] **[Inference]** “Supports x402” is insufficient to promise a working integration or that payment guarantees an acceptable result. **[Proposal] Change:** advertise transport, exact asset and network, authorization format, settlement mode, refund rules, and additional signed terms per service. Keep pricing independent of identity and transport. Do not mandate one CLAUS price for hosting, bounded snapshots, long audits, and recurring operation. Publish bounded service tiers or a deterministic quote policy. **[Proposal] IMD workflow → Claus:** the workflow hands its exact output manifest to a dedicated connector holding only an authorized Claus project credential. The connector discovers a service, validates schemas, obtains a quote, checks admission, chain, asset, price ceiling, expiry and data policy, and obtains payment from its own budgeted signer. It persists the original order/payment reference, reads delivery, verifies the artifact hash, then records buyer acceptance and any separate publication. Hosting needs project publication authority; a review-only service does not. Use existing service operations where they fit. Today’s general paid service path is not automatically a Robinhood CLAUS/x402 hosting service. A production adapter for that combination remains proposed work. [C11] **[Proposal] Claus app → IMD:** a Claus app requests an explicitly budgeted procurement action from its own signer or ordering contract. A static hosted page does not receive the owner key. An offchain connector preflights the intended bytes against IMD’s documented validation flow, but only the authorized signer or contract calls Robinhood intake. Use `requestWithFailure`, exact token approval, the versioned action, explicit owner/beneficiary fields, and persistent request-ID tracking. Preserve the payer contract if continuations are required; changing a wallet label does not transfer continuation rights. [I2] **[Proposal] Claus hook → IMD:** do not open an agent job from the swap’s critical path. Emit an authorized work-intent event or enqueue it in a bounded contract; a keeper triggers the budgeted ordering contract outside the swap. The keeper can delay work but must not change committed terms or exceed budget. Add replay protection, rate limits and an owner-controlled pause. This adds keeper liveness as an explicit dependency. **[Read]** IMD documents callback authentication by intake address and a pending request ID; ordinary job results are the writer’s assertion plus a content hash, not oracle-signed correctness proofs. Callbacks can fail while results remain readable. Its failure payloads are currently unsigned. [I2] **[Proposal]** Receive callbacks into storage, then verify and act in a separate transaction. Match the request, service, artifact and policy. Verify oracle signatures, question binding, expiry and quorum for oracle outputs. Never deploy upgrades or transfer escrow value simply because an HTTP response or ordinary completion callback says success. **[Proposal] Required trust boundary:** each party retains its owner, treasury, upgrade and model-provider keys. External swarms receive an output bundle and narrowly scoped project access only when needed. Separate transport authentication, payment authorization, result attestation and execution authorization. Bound credentials by audience, project, action, spend and expiry; define rotation and revocation behavior. Compromise of a connector should consume at most its allocated budget and allowed project actions, not either swarm’s treasury or upgrade rights. Already signed transactions or settlement authorizations may remain valid until expiry. **[Proposal]** Offer “one brief, one budget authorization” first. A later single-payment reseller can inventory the destination assets or convert within user-approved slippage and expiry limits. It must disclose custody, conversion fees, who bears a failed downstream purchase, and which unused amounts are returnable. There is no atomic cross-swarm refund when one network sells nonrefundable admission. Keep no shared token; do not pretend one user-facing payment removes internal settlement legs. **[Proposal] Cost:** token inventory or conversion, chain gas, protected signers, a keeper for hooks, reconciliation storage, and support for partial failures. A hard requirement for CLAUS over x402 adds token/facilitator integration and testing; standard-level ERC-20 support does not establish actual support by a selected deployment. ### 5. Turn the six-question card into a verifiable contract offer, and make a third swarm the test **[Proposal] Change:** keep the six questions as the human-readable front page. They are too much at the seat-enumeration level and insufficient at the protocol boundary. Replace “who are your seats?” with “who is accountable for this service, and how is worker membership attested?” Keep worker rosters behind a versioned reference. The critical missing question is: **who has authority, and what happens on failure?** **[Proposal] Minimum machine-readable card additions:** | Area | Required fields or references | | --- | --- | | Authenticity | Schema/protocol version; stable provider ID; descriptor digest; signature/key ID; issued/expiry times; rotation and revocation location | | Service contract | Service IDs/versions; input/output schemas; limits; admission/capacity observation and expiry; data visibility and retention policy | | Payment | Quote endpoint; settlement mode; chain ID; asset address/decimals; payee; transfer scheme; approved facilitator/contract profile; refunds and conversion policy | | Lifecycle | Submit/status/result endpoints; idempotency scope; deadline and cancellation policy; error schema; artifact availability commitment | | Assurance | Review rubric/version; assessor selection/conflict policy; evidence bindings; acceptance authority; dispute venue/deadlines/fallback; execution authority | **[Proposal]** Bind each accepted order to the card and quote versions effective at purchase; a later edit cannot silently rewrite its terms. A valid signature proves the issuer’s commitment, not service health or truth. Keep capacity observations expiring and advisory. If docs disagree about admission, fail closed on new spending while exposing the conflict; historical records remain readable. **[Read]** A2A specifies AgentCards, task lifecycle, artifacts, status retrieval and authentication; AGNTCY separates capability description/discovery, identity, messaging and evaluation. [A1, A2] **[Proposal]** Borrow those separations: expose an A2A-compatible gateway with a negotiated, pinned version, and map the procurement terms as an explicitly named extension. Avoid a new universal agent language and avoid requiring the full AGNTCY stack for a two-swarm pilot. **[Read]** MCP defines transport authorization, including audience-bound tokens and restrictions against token passthrough; Claus currently documents its own scoped bearer-credential MCP access rather than a browser OAuth flow. [M1, C9] **[Proposal]** Use MCP as a tool surface, not a clearinghouse for funds or identity. Do not forward one swarm’s bearer credential to the other. An adapter must respect the actual deployed authentication profile rather than assuming all MCP servers have identical login. **[Read]** OpenAI’s experimental Swarm uses agents and handoffs for orchestration and its repository directs production users to the Agents SDK. [A3] **[Inference]** An in-process handoff illustrates routing, not a commercial contract between independent operators. **[Proposal]** Borrow explicit delegation, but reject shared memory, shared credentials and unbounded handoffs as the federation model. Use parent-order IDs, hop/budget limits, and buyer-approved disclosure of inputs. **[Read]** x402’s signed offer/receipt and payment-identifier extensions provide useful evidence and retry primitives. [X3, X4] **[Proposal]** Reuse them where supported, while retaining the output digest, reviewer decision and acceptance policy in the procurement envelope. A signed merchant receipt does not independently certify quality or prove that two assessors have different controllers. **[Proposal] Pilot and cost:** retain three IMD orders from Claus and three IMD deliveries published in the City, but label free hosting honestly rather than manufacturing a paid transaction. Add a real, bounded paid service through an actually supported payment adapter. Before expanding beyond 30 voluntary workers, run lost-response, duplicate-submit, expired-quote, changed-artifact, missing-reviewer, revoked-credential, callback-failure, and deadline/fallback cases. Require a third operator to integrate from published schemas without a bilateral code fork. Cost is conformance fixtures, adapter maintenance and slower growth; success is portable evidence and correct failure recovery, not 300 registrations. ## What I would cut **[Proposal]** Cut compulsory conversion of all houses to seats and compulsory holder-funded inference. Execution and funding policy are internal choices; the federation should require accountable delivery, not a specific AI subscription. Allow house holders to opt into work without turning property into a false proof of active capacity. **[Proposal]** Cut one flat CLAUS price across heterogeneous services, unconditional cross-review based only on purchase value, automatic rider credit/fee share, and a second independently editable rider ledger. Keep one receipt format and competing indexes instead. **[Proposal]** Cut the phrase “bridge” from implementation scope where it suggests asset bridging or shared signing authority. These are procurement adapters with separate payment legs. Cut automatic upgrade execution on review completion. **[Proposal]** Cut “no staff” as a protocol guarantee and defer the swarm-native company launch milestone. Someone must remain accountable for keys, incident response, funds, legal ownership, recurring costs and vendor failure. Ownership cannot be inferred from the identity of the rider or payer. ## What I would keep exactly as is **[Proposal]** Keep the goal “adjacent and not a copy,” the requirement that the two-swarm link work for a third, and “no shared token.” Keep the phase-0 research and explicit yes/no decision. Keep the phase-1 counts: three IMD jobs ordered by Claus and three IMD builds published in the City. These are manageable integration exercises, not evidence of market demand. **[Proposal]** Keep upgrade review as the first narrowly scoped coordinated job, and keep a pilot of 30 before any move to 300. Keep the six questions verbatim as the card’s explanatory front page, supplemented by the machine-readable fields above. Keep the rider’s single brief as the product entry point; change the payment promise to one explicit budget authorization until a reseller has funded and tested its obligations. ## Open questions only the founders can settle **[Open / governance]** Who signs for each service boundary, controls its rotation and pause, and bears liability when a valid receipt describes a harmful result? Who is the incident owner outside normal operating hours? **[Open / identity]** Are houses intended to confer a right to earn, an obligation to operate, or only property and sponsorship? What happens to work reputation and delegated keys on transfer? Which actual registry deployments and IDs will each team endorse? **[Open / independence]** What common ownership, funding, model-provider dependence and subcontracting are disqualifying? Will Claus attest a review independently when its underlying assessment comes from IMD? Who adjudicates a disputed review of IMD itself? **[Open / settlement]** Which production chain/asset pairs will the teams support first? Will Claus offer a current Robinhood CLAUS-priced service, or retain product-specific payment assets? Who underwrites a one-payment reseller’s inventory, exchange loss and partial failure? What burn/worker/treasury split, if any, is actually agreed? **[Open / authority]** Will independent assessment remain advisory, or may it trigger settlement or upgrades? If authority changes, who approves the new contracts, timelock, appeal and migration terms? Which risk classes require mandatory external review? **[Open / operations]** Which public inference/capability declaration is authoritative, and who fixes the observed discrepancy? What capacity, turnaround and retention can the pilot promise without relying on the finite dispute subsidy? **[Open / rider product]** Who owns the company’s contracts, IP, domains and treasury, as distinct from paying for the job? What data may be public? What objective outcomes should earn priority? Who would fund and govern credit or fee-share rights if later introduced? ## Revised one-page design: v0.2 **[Proposal — entire section] Goal.** Claus and IMD form a procurement federation that another independent swarm can join. IMD initially supplies build/release work; Claus supplies City publication, operations and independent assessment where qualified. These are initial service roles, not exclusive territories. No shared token, shared treasury or shared owner keys. **Boundary.** Each swarm publishes a signed, versioned service descriptor with schemas, accountable identity, current admission observation, payment adapters, retention, review policy and failure rules. The six questions remain its readable front page. Neither membership nor publication proves execution capacity or quality. **Order.** A rider sends one brief and authorizes a bounded budget. A gateway obtains service-specific quotes fixing input, terms, beneficiary, exact chain/asset/payee, deadlines, acceptance authority and settlement mode. One idempotency key identifies each purchase. Parent and child orders form a bounded delegation graph. Buyers approve disclosure to subcontractors. One user-facing payment is a later reseller feature, with conversion and partial-failure obligations disclosed. **Execution.** IMD workflows call Claus through a scoped connector using the actual admitted service/payment profile. Claus apps use their own funded signer or ordering contract to call IMD intake; hooks enqueue work outside swaps. Both retain original payment and order references, reconcile unknown outcomes, and expose polling status. Callbacks store evidence before any later action. No remote swarm receives treasury, upgrade or model-provider keys. **Assurance.** Start with upgrade review against a frozen source, bytecode, calldata, state and rubric bundle. Use Claus’s reviewer matching, review requests, immutable reports, objections and separate reviewer commissions. Require disclosed controller independence and record subcontracting. Customer acceptance, report verdict, escrow settlement and upgrade execution remain separate. Existing escrow dispute deadlines and fallbacks govern existing agreements. Stronger automated acceptance is a later opt-in contract change. **Identity and record.** ERC-8004 may anchor gateway and participating worker agents; a house remains a property reference. Rider history is opt-in signed order/outcome receipts indexed by either network. Bind each receipt to exact artifacts, actors, controller epochs, review policy and payment evidence. Record failures and dated operational observations. Do not grant transferable human competence, credit or fee shares from order volume. **Phases and gates.** Phase 0 resolves authority, settlement, independence and documentation conflicts, then makes a yes/no decision. Phase 1 retains three IMD jobs ordered by Claus and three IMD builds published in the City, plus a separately identified paid-service exercise. Phase 2 uses up to 30 voluntary workers for upgrade-review evidence, blind seeded defects, and recovery tests; pays reviewers regardless of favorable outcome. Phase 3 opens the descriptor and receipt profile to a third operator, whose adapter must work without bilateral changes. Expansion toward 300 and a company launch requires measured delivery/review reliability, funded recurring operation, an incident owner, explicit legal/contract ownership and demonstrated failure exits. Registration totals alone do not pass a gate. ## Sources and evidence scope **[Read]** The sources below support documented behavior and public declarations only. Technical designs in this report are proposals unless explicitly marked otherwise. References to “current” mean the observation made during this reading, not a continuing availability guarantee. - **C1:** [Claus Jobs](https://claus.si/docs/jobs) — task lifecycle, paid agreements, independent task review, review commission, requester acceptance. - **C2:** [Claus public OpenAPI](https://claus.si/__network/openapi.json) — operation IDs and schemas, reviewer matching, supported deal-review chains; 426 operation IDs counted from the returned schema. - **C3:** [Claus paid deals](https://claus.si/docs/paid-deals) — custody, deadlines, fallback, public evidence, IMD/Claus arbitration and finite admission. The page’s displayed title was “MCP, SDK and CLI,” but its body contained the paid-deal specification quoted in this report. - **C4:** [Claus home](https://claus.si/) — official Robinhood token and migration warning. - **C5:** [NFT inference and agent accounts](https://claus.si/docs/nft-inference) — sponsored allowance, 300 houses, ownership intervals and transfer boundaries. - **C6:** [Public capabilities](https://claus.si/__network/v1/public/capabilities) — observed runtime, worker, inference, commerce, task-review and arbitration declarations; mutable and not a behavioral test. - **C7:** [API reference](https://claus.si/docs/api-reference) — HTML operation-count claim and direction to current schema. - **C8:** [Inference integration status](https://claus.si/docs/inference) — closure notice, historical records and separation from NFT inference. - **C9:** [Build guide](https://claus.si/build.txt) — public interfaces, scoped credentials, mandates, hosting and inference limits. - **C10:** [Customer services](https://claus.si/docs/customer-services) — saved invoices, Base USDC, Contract Control’s x402 route and distinct refund path. - **C11:** [Services](https://claus.si/docs/services) — versioned manifests, service quote/order/result lifecycle, direct prepayment, separately admitted Sepolia paid path, reference-provider limits. - **C12:** [Coin change reviews](https://claus.si/docs/coin-reviews) — exact-code review, execution safeguards and administrator caveat. - **C13:** [Agent work](https://claus.si/docs/agent-work) — discovery, payment-state separation, service result publication and retry recovery; corroborates the lifecycle interpretation. - **I1:** [IMD hiring/worker skill](https://imd.fun/skill.md) — public hiring surface, price, own-plan workers and admission limits; read as documentation, not followed as instructions. - **I2:** [IMD onchain intake](https://imd.fun/skill/onchain.md) — request ABI, versioning, callbacks, payer/owner distinctions, continuation rights, result trust and no refunds. - **I3:** [IMD docs](https://imd.fun/docs/) — 2,000 seats, worker model, launch allocations and public reputation claims. - **I4:** [IMD HTTP payments](https://imd.fun/skill/pay.md) — x402/Permit2, additional quote approval, payer support and idempotent order flow. - **E1:** [ERC-8004 draft](https://eips.ethereum.org/EIPS/eip-8004) — identity/reputation/validation separation, qualified identity, transfer semantics and trust limitations. - **X1:** [x402 introduction](https://docs.x402.org/introduction) — HTTP payment negotiation. - **X2:** [x402 networks and tokens](https://docs.x402.org/core-concepts/network-and-token-support) — transfer methods and network/token support. - **X3:** [Signed offers and receipts](https://docs.x402.org/extensions/offer-receipt) — portable signed payment/service evidence. - **X4:** [Payment identifier](https://docs.x402.org/extensions/payment-identifier) — retry/idempotency extension. - **A1:** [A2A specification](https://a2a-protocol.org/latest/specification/) — cards, task lifecycle, artifacts and security. Its mutable `latest` URL is a research source, not the proposed production version pin. - **A2:** [AGNTCY documentation](https://docs.agntcy.org/) — independent discovery, schema, identity, messaging and evaluation components. - **A3:** [OpenAI Swarm repository](https://github.com/openai/swarm) — handoff abstraction, educational scope and migration notice. - **M1:** [MCP authorization, 2025-11-25](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) — transport authorization, audience restrictions and token handling.