# Claus beside IMD: federate work contracts before federating seats Independent view 2 — protocol and interface lens. Public-source desk research, 12 November 2026. **[Proposal]** Proceed with a narrow federation experiment, but do not adopt v0.1 as the protocol. My five changes, in order, are: define a recoverable work contract; reuse Claus's existing review and escrow machinery; install restricted adapters rather than a privileged bridge; separate agent, property and rider identity; and replace the universal CLAUS price and premature rider incentives with explicit service economics. The scarce interface is the agreement between organizations, not a common implementation of their workers. ## Evidence conventions and limits **[Read]** means the cited public source states the claim; it does not mean I independently verified its implementation or deployment. **[Inference]** means a conclusion from the cited evidence. **[Proposal]** means my recommendation, including suggested fields, policies and experiments. These labels apply to the whole paragraph, list item or table cell unless another label appears. Source references resolve to the bibliography below; short quotations preserve the source's wording. **[Research limit]** I read the public pages, machine-readable descriptions and public repositories cited below. I did not log in, purchase anything, submit jobs, execute contracts, test private APIs, inspect other researchers' views, or establish real-world reviewer independence. Public GET observations are dated to this research session; advertised readiness is not a successful transaction. The supplied description of Matt Huang's November 12 post is a premise of the brief: I have not independently located and verified that post and do not attribute additional claims to it. ## Correct the baseline first | v0.1 premise | What the public evidence supports | Consequence | | --- | --- | --- | | Claus has no coordinator or independent acceptance. | **[Read]** Claus documents `workflow_create`, milestone briefs, work profiles, reviewer matching and evidence-bound task reviews. Its workflow profile is specifically `recurring-public-free-tasks-v1`; acceptance remains the requester's. The review schema says: “Does not accept the task or release funds.” [C1–C4] | **[Inference]** There is already limited coordination and an advisory review protocol. The missing feature is an agreed acceptance/authorization policy across swarms, not an empty coordinator slot. | | Claus has no onchain record. | **[Read]** Jobs can bind task commitments to ETH or CLAUS escrows; public payment reads identify the observed chain state. Separately, Claus publishes some Journal records onchain. A Journal record is not independent review, and neither establishes a general seat reputation registry. [C3, C5, C10] | **[Proposal]** Say “no demonstrated common cross-swarm work/review receipt,” rather than “no onchain record.” Reuse existing commitments. | | Claus has no posted price; its x402 services use USDG. | **[Read]** Work profiles can quote ETH or CLAUS. Customer-services documentation lists Contract Control initially at 0.25 USDC per address and a one-shot x402 route paid in **Base USDC**. The generic service-manifest guide separately describes a paid wallet/testnet path priced in Sepolia wei and direct prepayment. These are different products. [C2, C6, C7] | **[Inference]** USDG is not supported by the current customer-service documentation I read. Nor does a priced observation service establish a priced upgrade-review or hosting service. | | New project inference is off, so Claus needs holder-funded seats. | **[Read]** The inference guide says, “New project inference requests are disabled as of 21 October 2026.” However, the public capabilities response read in this session has `inference.enabled: true`, and product-reference reports project inference `enabled`. NFT sponsorship is separately configured, up to $5 per NFT per day, without automatic top-up. [C1, C8, C9, C11] | **[Inference]** Current legacy-inference admission is unresolved because first-party sources conflict. No inference submission was attempted. NFT, native-resident and legacy-project budgets must not be collapsed into one availability assertion. | | The current pool fee funds half burn and half compute/hosting. | **[Read]** The migration guide and configuration reference describe Robinhood's 2% project swap fee as 1% buyback/burn plus 1% operating treasury. The discovery text in `llms.txt` instead says all Robinhood project fees go to liquidity, with no treasury share. The configuration reference gives a broader operating purpose: inference, compute, hosting and ongoing work. [C10, C12, C13] | **[Proposal]** Retain the guide/reference allocation as a published description, explicitly flag the conflicting discovery text, and do not treat a treasury purpose as a guaranteed service budget or a service-request fee split. This report does not certify live fee routing. | | There are 395 API operations; houses are prospective seats. | **[Read]** The fetched OpenAPI identifies version `0.41.0`; counting its GET/POST/PUT/PATCH/DELETE operations gives 426. NFT docs describe 300 numbered properties, owner-bound account access and ownership intervals; transfer does not transfer private projects, history or keys. [C4, C9] | **[Inference]** Neither operation count nor house count measures functioning, independent workers. Houses are access/property objects, not automatically task executors. | | IMD has a ready general-purpose workflow door into Claus. | **[Read]** IMD's documented `workflow.open` performs contracts, review, deployment, frontend, IPFS/ENS publication and validation. Its draft is strict about unknown fields and specifies IPFS hosting. A native paid Claus workflow step is not documented there. IMD's live capabilities advertise HTTP payment on Ethereum and a Robinhood intake; those are separate payment surfaces. [I1–I4] | **[Proposal]** An external adapter can follow IMD and order a Claus service today where admitted. A native external-service step would be new IMD work; do not imply that changing a brief adds that capability. | **[Read]** IMD's seat documentation supports the brief's 2,000-seat model, holder-supplied machine/AI plan, and 8% connected-seat plus 2% accepted-work launch allocation. Its API documentation also explains the limit of an accepted record: for text outputs the verifier checks structure alone; `acceptance-v2` records accepted/rejected results and categories, including rejections. [I5, I6] **[Inference]** Thus “accepted,” “independently checked,” “paid,” “deployed” and “still useful” are different predicates on both networks. A cross-swarm protocol that puts them in one success score would discard information the source systems already distinguish. [C3, C5, I3, I6] ## Five changes, ranked ### 1. Standardize a versioned work contract and receipts, not just a swarm card **[Proposal — change]** Make the unit of interoperability a service-specific agreement and its immutable evidence, carried over ordinary HTTPS. Each swarm keeps its own coordinator, seats and token. Begin with a small adapter profile: discover capabilities, obtain a quote, submit an authorized order, read status/result, and append a review or acceptance receipt. Notifications and cancellation are optional advertised capabilities; polling and explicit expiry are mandatory. **[Read — reason]** A2A already supplies agent discovery, task IDs, artifacts, asynchronous task handling and authentication declarations. Its idempotency section says, “Send Message operations MAY be idempotent.” x402 supplies payment requirements and settlement messages; ERC-8004 explicitly says, “Payments are orthogonal to this protocol and not covered here.” None of these statements defines a bilateral service's acceptance criteria, refund policy or deployment authority. [S1–S3] **[Inference — reason]** The six-question card is useful prose but insufficient as an executable interface. “Who checked it?” is a per-result fact, not a static property a provider can promise by naming itself. A third swarm should implement one profile, rather than reverse engineer the IMD–Claus relationship. [C3, S1–S3] **[Proposal — smallest common objects]** Use the following five objects. These are new interface concepts, not claimed existing endpoint names: | Object | Required contents | | --- | --- | | Capability/card | Schema and protocol versions; swarm identity; service/version; endpoint and authentication method; input/output schema references and digests; current admission/limits; quote/status locations; supported payment and review profiles; operator/signing-key reference and rotation policy. Seat roster is optional internal provenance. | | Quote/agreement | Buyer and provider; buyer-generated idempotency key; service and immutable version; exact input and criteria digests; quote ID/expiry; asset contract, chain, decimals, integer amount and recipient; admission-only/prepayment/escrow semantics; delivery/review deadlines; failure/refund/cancellation terms; selected reviewer/arbiter and exact authority; permitted onward disclosure. | | Order/status | Globally namespaced order ID; parent workflow/child order links; accepted agreement digest; native task/intake identifiers; payment evidence; separate execution, review, buyer-acceptance and settlement states; next actor; deadline; monotonic revision and error/recovery code. | | Result manifest | Order/agreement binding; artifact URI, media type, byte length and algorithm-tagged digest; source commit, build and test evidence; chain/address/block references where relevant; declared limitations; publication authority and retention terms. | | Review/acceptance receipt | Receipt type and issuer; order/agreement, exact result and rubric digests; per-criterion pass/fail/insufficient-evidence; methods and evidence; reviewer/operator conflicts; timestamp/expiry; issuer authentication; objections and superseding receipt links. Buyer acceptance is a distinct receipt. | **[Proposal — card verdict]** Keep the six questions as the human introduction. Add two indispensable questions: **“Where is the versioned machine interface?”** and **“Who can authorize what, and what happens when it fails?”** Replace mandatory disclosure of every seat with a provider identity and a provenance policy. Link changing prices/admission to an expiring quote endpoint, not a forever-current price in a cached file. Advertise the review policy on the card; attach the actual review to the result. **[Proposal — recovery contract]** Retrying an order key with identical agreement/input returns the original order; different input under that key is a conflict. Persist original transaction hash, nonce, quote and native ID before dispatch. An unknown payment stays `unknown`, never `unpaid`; read/reconcile it before authorizing another purchase. At-least-once events must not cause duplicate jobs or payments. Separate `delivered`, `review_passed`, `buyer_accepted`, `escrow_credited` and `withdrawn`; expose `blocked`, `expired`, `disputed` and `insufficient_evidence` without rewriting history. Cancellation stops only effects the native service permits; it cannot reverse spent IMD or erase escrow obligations. [C3, C5, I1, I3 provide the motivating native semantics.] **[Proposal — cost]** Two adapters, a small schema/profile document, durable deduplication storage, receipt export and compatibility fixtures. This delays the attractive single-screen demo and forces agreement on edge cases. Keep A2A optional at first: borrowing its lifecycle vocabulary is cheaper than implementing every transport. Specify and pin the selected A2A version if adopted; the page read here identifies latest release 1.0.0, so do not silently mix it with 0.3 examples. [S1] ### 2. Reuse Claus's existing reviews and deals; make independent verification distinct from acceptance **[Proposal — change]** Start with an **upgrade-review service**, not automatic acceptance of upgrades. Commit the reviewer, measurable rubric, target commit, current/proposed implementations, chain state and proposed calldata before work begins. A second run is one verification method; it is not the definition of independence. Record common ownership and shared model/tool/provider dependencies. Require a fresh review if any reviewed artifact or calldata changes. **[Read — existing sequence]** Claus already documents the essential sequence: select a reviewer through matching; commit `reviewPolicy` before claim; deliver and, for paid work, submit matching onchain delivery; freeze evidence with `app_task_review_request`; submit one immutable per-criterion report with `app_task_review_submit`; optionally object with `app_task_review_challenge`. A paid reviewer gets a separate linked commission through `app_task_review_commission` and the normal funded Jobs flow. The Jobs guide says, “A URL is attribution, not a frozen copy of a remote page,” and “Review is advisory.” [C2–C5] **[Proposal — use it]** Adapt that sequence instead of adding a competing review database. Freeze actual source/evidence bytes and their hashes, not merely branch names or URLs. Pay for a properly completed report even if its verdict is fail or insufficient evidence. Publish the reviewer selection basis and known conflicts. An IMD-built artifact reviewed by Claus should carry both the original IMD provenance and the exact Claus report binding, rather than acquire a generic “cross reviewed” badge. **[Read — settlement and arbitration]** Claus's deal guide gives a 24-hour delivery review window; silence then pays the provider. A dispute has three days; without a valid resolution the buyer gets a refund. Its supported public-deliverable dispute path asks an IMD oracle panel of five, requiring four identical payout answers with zero numerical tolerance, then requires a separate Claus decision on that exact split. The resolution is an expiring agreement-specific signature which a party must submit, not an automatic payment. At observation, capabilities advertised arbitration enabled, a healthy worker and seven remaining cases from a finite eight-case allocation. [C1, C5] **[Inference — reason]** Existing task review, payment arbitration and privileged upgrade approval are three different protocols. A dispute path that already depends on IMD cannot count as independent of IMD for every conflict involving IMD. Five seats can also share tools or operators; consensus alone does not establish diversity. Deadline-bound escrow cannot wait indefinitely for an uncertain cross-swarm review. [C3, C5, I6] **[Proposal — acceptance policy]** For the pilot, keep requester acceptance and owner-controlled upgrade execution. The independent reviewer attests the frozen evidence; a requester-side policy checks that the required review exists and no agreed blocker remains before accepting or deploying. This policy gate is new integration work, not a current Claus escrow feature. Never automatically deploy from a positive text verdict. If a binding automated upgrade gate is later desired, put the reviewed commit/build/calldata commitments and signer/quorum policy in a separately reviewed execution module. Until then label bypassable checks advisory. **[Proposal — deadline policy]** Admit a paid review only when its maximum turnaround plus an explicit buyer response margin fits inside the original escrow window. If evidence is missing or a reviewer disappears, record inconclusive/expired and alert the buyer early enough to dispute onchain. A public objection does not pause the clock. Choose a neutral pre-agreed fallback for disputes where IMD or Claus is conflicted; do not promise the existing subsidized path is unlimited insurance. [C3–C5] **[Proposal — cost]** Reviewer compensation independent of verdict, more frozen evidence storage, conflict screening, a deadline monitor and a buyer policy agent. Binding deployment prevention would require additional contract work and review. Some requests must be rejected rather than forced through a cheap, flat-price rubric. ### 3. Replace “bridge both ways” with two restricted adapters and an explicit trust map **[Proposal — change]** Implement a durable workflow adapter outside ordinary build workers. It holds scoped project credentials and calls a separate budget-enforcing signer. It never shares rider wallet keys or token-owner keys with either swarm. Route work receipts, not arbitrary code or arbitrary signing instructions. Preserve the native payer across project continuations. **[Read — IMD to Claus constraints]** IMD's documented workflow already includes IPFS/ENS publication; it does not document a paid Claus step. Claus's current customer-services product offers Contract Control via a Base-USDC x402 GET, but that service is a standard-slot/owner observation, explicitly not a security audit. Generic Claus service composition has immutable versions, schemas, leases, quote/accept/result operations and direct-prepayment semantics; its reference providers are Claus-operated, not independent adoption. Hosting still requires a separate publication action. [I2, C6, C7] **[Proposal — clean IMD to Claus path]** Initially let the adapter observe a completed, hash-checked IMD delivery and request an explicitly admitted Claus service under a separate child order. For a paid review or hosting product, first publish the actual service contract, price, input/output schema, funding route and retention terms; these are not supplied by the existing Contract Control offer. Until that exists, publish with ordinary scoped Claus publication credentials and commission review through Jobs. Calling a free publication API should not be represented as a paid hosting purchase. A native IMD external-service node can follow later, with an allowlisted service/version, bounded budget and mandatory receipt; the current strict workflow schema must be extended deliberately. [C2–C7, I2] **[Proposal — x402 handling]** At a paid HTTP boundary, check the advertised version, scheme, chain, exact asset/recipient/amount and expiry before the signer acts. Bind payment to the accepted input/quote where the provider supports it. Persist settlement proof and the returned artifact digest; an HTTP 200 or payment authorization is not itself proof of acceptable work. Honor native unknown-settlement reconciliation rather than buying the same result again. For IMD's HTTP route specifically, implement its extra EIP-712 quote approval: the documentation says a stock x402 client produces the payment signature but not this second signature. [S3, I1, C6] **[Read — Claus to IMD door]** IMD's Robinhood intake is advertised at `0xa43e6f75ee006411f79ac1c84120606c2330de82` on chain 4663; its asset is IMD `0x5f7bb59365ce557c26dbcaa4ee9d39a4b95b7127`, with a 0.5 IMD job price. The interface includes `requestWithFailure(bytes32,bytes,(address,bytes4),bytes4,address,uint256)`. Requests include public JSON and versioned action IDs. Payments are not refundable. The direct contract route itself does not require an HTTP quote, though IMD recommends checking and quoting exact input beforehand. [I3, I4] **[Proposal — clean Claus to IMD path]** A Claus frontend asks its user's wallet, or a purpose-built owner-controlled purchasing contract, to approve a bounded amount of chain-specific IMD and call that intake. Validate the body offchain before spending. A hook should enqueue/record a request through a constrained purchasing component with a budget and rate limit; it should never expect an HTTP service or delayed result to complete inside a swap. Return promptly, retain the native request ID, then let the adapter follow the asynchronous completion. Carry the original rider attribution separately from the paying contract. Set the intended contract owner and allocation recipient explicitly, subject to the launch kind's schema. **[Read — hidden payer consequence]** IMD says, “The address that called `request`” is the payer, “Not `tx.origin`.” Continuation belongs to that payer and cannot be changed; omitted launch ownership/allocation fields can default to it. The docs also say a job result is “Not signed by the oracle” and its meaning is “the writer's word.” Callback delivery is once, may revert, and can be late or missing. [I3] **[Proposal — callback handling]** Authenticate the intake and a pending request this buyer actually created; settle each ID once. Pin the expected intake per request so rotating a configuration address does not invalidate in-flight orders. Store the result reference, then retrieve and verify bytes and perform consequential actions separately. Handle refusal/no-result and callback loss through public status reads. Never turn a callback into a generic token-owner executor. If using an oracle attestation, additionally pin the question hash, consumer/domain, quorum and freshness; the intake request ID and oracle request ID are different namespaces. [I3] | Boundary | Trust that remains, based on the sources | Proposed restriction | | --- | --- | --- | | IMD job completion | **[Read]** Control-plane writer chooses the reported job result; a hash fixes bytes, not their truth. [I3] | **[Proposal]** Verify artifact bindings and independent review before accepting a consequential result. | | IMD oracle | **[Read]** Panel honesty, attester key and writer liveness remain assumptions. [I3] | **[Proposal]** Pin question/consumer and signer policy; expiry and value-at-risk limits; no automatic privileged action from an unrelated answer. | | Claus project API | **[Read]** Scoped credentials and current project permissions control API changes; owner-granted team roles do not imply financial/admin powers. [C2] | **[Proposal]** Give the adapter only needed publication/task scopes; signing is a distinct capability with separate limits. | | Claus arbitration | **[Read]** IMD review plus Claus approval yields a bounded resolution signature; availability and admission can change. [C1, C5] | **[Proposal]** Pin escrow/parties/amount/expiry; disclose the Claus decision-maker and conflict; retain direct exits. | | CLAUS on Robinhood | **[Read]** Configuration reference discloses a single developer admin, upgrade/mint authority and no enforced onchain timelock. [C13] | **[Proposal]** Cap exposure; monitoring is advisory unless the owner adopts an enforceable delay. A reviewer does not become an admin. | | Rider/adapter signer | **[Inference]** One top-level payment requires someone to hold or source the assets for separate downstream purchases. [C6, I1, I3] | **[Proposal]** Disclose custody, conversion and fee policy; allowlist recipients/actions; cap amount, lifetime and number of child orders. No automatic onward delegation. | **[Proposal — cost]** A signer boundary, durable event/reconciliation worker, transaction monitoring and integration security review. Mainnet-to-mainnet service purchases can need balances and gas on more than one chain. This is operationally less magical than “one payment,” but avoids making a shared hot key the federation's real protocol. ### 4. Use ERC-8004 for agent discovery/provenance; do not turn houses or riders into the same identity **[Read — reason]** ERC-8004 defines a global agent identity as registry namespace/chain/address plus `agentId`, an ERC-721 identifier resolving to a registration file. That file advertises services and supported trust models. The reserved `agentWallet` requires proof of control to change; it is cleared on transfer. Its security section states, “Sybil attacks are possible, inflating the reputation of fake agents.” Clients posting feedback need not themselves register. IMD separately binds a seat NFT and device to an ERC-8004 agent and writes work feedback through its WorkRegistry. [S2, I6] **[Proposal — what to register]** Register one agent identity for each externally accountable swarm service/coordinator, and optionally one for each **actual enrolled operator agent** whose work is attributed publicly. A Claus registration file should identify the service endpoints, supported protocol versions, scoped operational identity and payment wallet, and link the manifest. Record a house entitlement as a chain/collection/token/ownership-interval relationship to that operator, not as proof of a live machine or as an identity substitution. The ERC-8004 token ID need not equal the house ID. Do not register 300 fictional workers to make the seat count look complete. **[Proposal — control and transfer]** Keep property rights, agent registration ownership, operational session keys and payout authority separate. Publish the owner-authorized binding and revocation rules. On house transfer, expire its old entitlement and credentials; retain old work receipts under the operator and ownership interval that performed them. Do not export private history to a house buyer. If the operator also transfers its ERC-8004 identity, show the control-change boundary in reputation views; buying the identifier must not imply buying the prior operator's skill. [C9, S2 motivate these distinctions.] **[Proposal — rider record]** Identify a rider initially through its wallet/account and explicit signatures/delegations, not a mandatory transferable ERC-8004 NFT. A rider may appoint a registered agent, but the human/principal, payer, owning entity and execution agent are separate roles. Collect attestations referencing the same order: who commissioned, what was accepted, who delivered, who reviewed, what was paid and what was observed later. Each issuer signs its own assertion; neither swarm maintains an authoritative copy of the other's facts. Build both displays from these receipts and show disagreements. **[Proposal — reputation]** Attribute build quality to builders/reviewers and brief/acceptance behavior to riders; do not award a rider credit for simply spending. A “still live” observation needs the observed release, time, checker and measurement method; absence of a later check is unknown, not failure. Filter ERC-8004 reputation by trusted feedback issuers and service/category. Keep subsidy-funded and common-controller work visible as such. Export selected provenance to the reputation registry; use validation requests/responses only when there is an actual validator mechanism and supported deployment, not as a cosmetic seal. [S2, I6] **[Proposal — cost]** Identity mapping, ownership-interval indexing, revocation/rotation handling and gas for selected registrations/anchors. There will be fewer marketable “seats” and no universal scalar reputation. In return, a third swarm need not reproduce Claus's NFT access model or IMD's device enrollment. ### 5. Quote service economics separately; defer rider priority, credit and fee share **[Proposal — change]** Keep separate tokens, but make accepted assets a service capability, not a requirement that every Claus endpoint change to CLAUS. Retain working payment rails during the pilot. Quote upgrade review by a bounded class—chain, contract count, diff size, privileges/value exposed and evidence available—with an explicit maximum scope. Hosting/operation is a separately priced duration and resource budget, not an indefinite obligation bought by one admission fee. **[Read — reason]** Current documented rails differ: IMD HTTP uses Ethereum IMD/Permit2 and a quote signature; its Robinhood intake takes IMD directly. Claus Jobs supports ETH/CLAUS escrow, customer observations use Base USDC, and generic service documentation describes separate paid admission. IMD explicitly sells admission rather than a guaranteed result, with no refunds. [I1, I3, I4, C5–C7] **[Inference — reason]** A universal “one CLAUS price over x402” would require new token/payment support, service underwriting and conversion policy; x402 alone does not supply them. Nor does the pool's 2% fee determine the margin on a service. A flat “every upgrade” review product invites unbounded work and adverse selection. [S3, C6, C7, C13] **[Proposal — one payment semantics]** For the pilot, show one brief and an itemized authorization covering native child purchases; do not promise one onchain transaction. If founders later require one actual payment, make the router an explicit merchant: identify the custodian, quote FX/slippage and gas, fund every downstream leg, and define refunds for unspent funds versus spent admission charges. No shared token or token bridge is necessary. Any burn/seat/treasury split applies to the new service's revenue under published terms, separately from existing swap-fee policy. **[Proposal — rider incentives]** Start with an opt-in receipt history and measured outcomes. Defer preferential scheduling, credit and fee-sharing until the founders define evidence requirements, common-controller exclusions and who funds losses. A fee share for a rider who also selects reviewers can reward friendly acceptance or wash orders. Record such affiliations before introducing rewards. Do not infer borrowing capacity from accepted-job count. **[Proposal — cost]** Less token uniformity, quote maintenance, an explicit funded operating budget and more honest checkout details. A true single-payment merchant adds custody/conversion/accounting risk. Delaying rider incentives loses a growth story but avoids encoding a manipulable success metric before any inter-swarm service reliability is measured. ## What to borrow from other systems—and what to avoid | System/source | Read from source | My inference and proposed borrowing | | --- | --- | --- | | A2A specification | **[Read]** Agent Cards advertise interfaces, capabilities, skills and security requirements. Tasks have distinct submitted/working/completed/failed/canceled/input-required/auth-required/rejected states; artifacts carry outputs. Card signatures use canonicalization and JWS. [S1] | **[Proposal]** Borrow discovery, asynchronous task handles, artifact separation and version negotiation. Add required commercial idempotency and receipt bindings. **[Inference]** “Task completed” is not buyer acceptance or escrow settlement; A2A messaging interoperability does not itself resolve those financial predicates. | | MCP tools | **[Read]** Tools declare input schemas and may declare output schemas; results can contain structured content and execution errors. Tool annotations must be treated as untrusted unless from trusted servers. [S4] | **[Proposal]** Expose the adapter as a few scoped tools. Avoid making a tool description or an LLM's successful tool call a grant of wallet, publication or upgrade authority. | | Virtuals ACP Node SDK v2 | **[Read]** `JobSession` distinguishes client/provider/evaluator roles and open/budget-set/funded/submitted/completed/rejected/expired states. Actions are gated by role and status; offerings validate requirements and set expiry. The quickstart sets the buyer's own address as evaluator. [S5] | **[Proposal]** Borrow explicit commercial roles, versioned offerings, expiry and role-limited operations. **[Inference]** An “evaluator” field is not proof of independence—the quickstart demonstrates why. Do not copy SDK signer provisioning or registry dependencies as mandatory membership requirements for every future swarm. | | AutoGen Swarm implementation | **[Read]** The manager selects the next participant from handoff messages; the source documents that without a termination condition the group can run indefinitely. [S6] | **[Proposal]** Borrow explicit handoff targets, bounded turns and resumable state within a coordinator. **[Inference]** A conversational handoff between configured participants is not an inter-organization purchase contract. Do not export a shared transcript or unlimited agent loop as the federation interface. | | x402 and ERC-8004 | **[Read]** x402 describes payment requirements and verification/settlement; ERC-8004 supplies identity, reputation and validation registries while leaving payments orthogonal. [S2, S3] | **[Proposal]** Compose these layers rather than invent replacements. Keep receipts for the work itself. Avoid claiming any registered endpoint, paid response or onchain hash is an independent correctness certificate. | **[Research limit]** These are documented mechanisms and public code, not measured claims that these systems have solved reliable decentralized swarm commerce at scale. ACP and AutoGen were read as public repository documentation/source; no SDK was installed or executed. [S5, S6] ## What I would cut - **[Proposal]** Cut converting all 300 houses into holder-funded workers as a prerequisite. Admit a small number of demonstrably available operator agents and retain existing sponsorship as a separately disclosed option. Whether to copy IMD's worker economics is a product decision, not an interoperability requirement. - **[Proposal]** Cut a universal CLAUS price and obligatory CLAUS x402 migration. Start with the existing route appropriate to each service and publish exact per-order terms. - **[Proposal]** Cut automatic review of “every upgrade” and a value-only threshold. Use a narrow, quotable review scope; privileges, irreversibility and exposure matter even for a currently small token. Never confuse monitoring with an enforced upgrade delay. - **[Proposal]** Cut the jointly maintained authoritative rider score and immediate priority/credit/fee-share promises. Start with attributable receipts and separate observations. - **[Proposal]** Cut “no staff” as a launch guarantee and “swarm native company” as the first interoperability milestone. A launch does not settle ownership, emergency response, bills, liability or ongoing operations. Require a named accountable owner and bounded operating mandate before that experiment. - **[Proposal]** Cut expansion to 300 seats as the definition of success. Require interoperability with a third independently operated adapter and successful recovery from failure instead. ## What I would keep exactly as is - **[Proposal — keep]** “Adjacent and not a copy,” and the requirement that the two-swarm link work for a third. Independence of internal coordination is the design constraint. - **[Proposal — keep]** “No shared token.” Each swarm retains its economics; services declare what their existing payment routes accept. - **[Proposal — keep]** Upgrade review as Claus's first coordinated job type, with IMD building and Claus housing, operating and reviewing as the initial product focus. Keep this focus without claiming an exclusive technical boundary: both already describe overlapping functions. [C1–C7, I2, I5] - **[Proposal — keep]** Phase 0 research and an explicit yes/no decision. Keep Phase 1's exact three IMD jobs ordered by Claus and three IMD builds published in the City as the bounded demonstration, subject to normal budgets and permission. Count publication separately from paid hosting and independent acceptance. - **[Proposal — keep]** One brief as the rider-facing product, and a public work record. Keep the six discovery questions verbatim as the human-facing introduction, while adding the machine and failure/authority fields above. Do not keep stronger claims about one-payment atomicity or reward eligibility. ## Open questions only the founders can decide 1. **[Proposal — decision needed]** Who is the seller to the rider: one accountable merchant underwriting both swarms, or a router arranging independent purchases? Who absorbs a spent IMD admission fee when the downstream review or hosting fails? 2. **[Proposal — decision needed]** Who funds pilot reviewer commissions, adapter gas, custody/conversion and ongoing hosting? What are the per-order and total-loss ceilings, and which accepted assets must the pilot actually support? 3. **[Proposal — decision needed]** Is independent review advisory, required by a requester policy, or a binding condition on upgrade execution? Are Claus's contract owners willing to adopt an enforced delay/gate, and who holds emergency powers? 4. **[Proposal — decision needed]** Who is eligible as an independent reviewer, which affiliations must be disclosed, and who arbitrates when Claus or IMD is a party? Will that neutral party be paid and available within existing escrow deadlines? 5. **[Proposal — decision needed]** What does a rider legally and operationally own: token balances, contracts' owner roles, IP, domains, revenue rights, or an entity? Who is the durable native payer for continuation, and can the rider replace the router without losing project control? 6. **[Proposal — decision needed]** Which evidence may be public, how long must results remain available, and who pays to pin them? What private-data workload, if any, should be excluded from the public-evidence pilot? 7. **[Proposal — decision needed]** Are house-linked workers optional, and what rights should survive house transfer? Which metrics would justify rider priority, credit or fee shares, and who bears default losses? 8. **[Proposal — decision needed]** What concrete reliability and margin targets warrant expansion? I recommend agreement on thresholds before the first pilot purchase, rather than changing them after seeing results. **[Uncertainty, not a founder preference]** The conflicting inference flags, stale discovery fee description and dated token-control observations need reconciled public evidence from their maintainers. Those are factual publication issues, not matters to resolve by choosing whichever answer favors this proposal. Mainnet paid review/hosting availability and independent operator diversity remain unproved in this desk review. [C1, C5–C13] ## Revised one-page design: work federation v0.2 **[Proposal — goal]** Make Claus the first independent operating/review partner beside IMD. One rider brief produces traceable purchases and results across independent swarms. Each retains its own coordinator, membership and token; a third swarm joins by implementing the same small service profile. **[Proposal — first product]** IMD produces a pinned build; Claus publishes an approved static release and coordinates a separately commissioned upgrade-review report. Begin with a bounded review class and observable criteria. The report is not authority to deploy. The owner retains contract keys and accepts execution risk under a separately declared policy. **[Proposal — protocol]** Publish a versioned swarm/service card answering the original six questions plus interface and failure/authority questions. The shared objects are capability, quote/agreement, order/status, result manifest and review/acceptance receipt. Pin service version, exact input/output/criteria hashes, parties, deadlines, price and settlement semantics. Separate execution, review, acceptance and payment states. Mandatory stable IDs and reconciliation prevent event retries from buying work twice; uncertain outcomes remain unknown. **[Proposal — transport and payments]** Start with HTTPS adapters and public status reads, using native IMD intake and Claus APIs. Optional A2A/MCP façades expose the same profile. Claus apps buy IMD through the Robinhood intake using a user wallet or capped purchasing contract. An external adapter follows IMD delivery before making the next admitted Claus purchase/publication. Each child purchase uses its documented asset and chain. One brief and one itemized authorization come first; one literal payment requires an explicitly funded merchant/conversion policy. No shared token or token bridge. **[Proposal — trust and keys]** Each swarm signs or authenticates only its own claims. Restricted project credentials do not grant wallet powers. A separate signer allowlists actions, recipients, budgets and expiry. Callback consumers authenticate their own pending requests and verify result hashes; consequential actions happen later. Pin trust/configuration per order, support rotation for new work and preserve old receipts. Never share rider or upgrade keys. **[Proposal — identity]** Use ERC-8004 handles for accountable service agents and enrolled workers, with manifests and explicit entitlement mappings. Houses remain property/access rights; transfer ends the old interval without moving private history. A rider is an accountable principal with a receipt history and optional delegated agents, not automatically another transferable seat. Feedback identifies its issuer, work category and exact evidence; no combined credit score yet. **[Proposal — review and disputes]** Reuse Claus's precommitted task review policy, frozen evidence, selected reviewer, immutable criterion report and separate review commission. Buyer acceptance, payment settlement and privileged deployment remain distinct. Set review deadlines inside native escrow deadlines and alert before dispute cutoffs. Disclose common controllers and shared dependencies. Use existing arbitration only where admitted, funded, timely and unconflicted; specify a neutral fallback rather than promising universal adjudication. **[Proposal — phases and gates]** Phase 0: reconcile public capability conflicts and agree prices, authority, privacy and risk limits; then yes/no. Phase 1: three Claus-ordered IMD jobs and three City publications, each with original IDs and receipts. Phase 2: admit up to 30 actual operators for the bounded upgrade-review pilot, expanding only as demand requires. Exercise reviewer timeout, negative/inconclusive verdict, duplicate notification, unknown payment, callback loss, revoked key, changed artifact and disputed settlement. Phase 3: a third independently operated adapter completes the same protocol without a custom bilateral key or identity migration. Phase 4: scale only after agreed reliability, turnaround, cost and outcome thresholds; consider rider incentives and an owner-accountable company experiment separately. Seat count is capacity, not proof of interoperability. ## Sources All sources below were retrieved/read on 12 November 2026. They are public primary documentation, machine-readable first-party data, standards or public source code. Living URLs can change; quotations and observed values above identify what was read. No citation is a claim that a paid flow was executed. - **C1** — [Claus public capabilities](https://claus.si/__network/v1/public/capabilities). Observed schema `1.0`, policy `site-capacity-2026-10-14-v1`; collaboration, review, deal/arbitration, workflow and inference fields. Volatile availability flags and case counts are session observations. - **C2** — [Work in Claus Network](https://claus.si/docs/agent-economy). Work profiles, asset-specific quotes, matching, portfolio, review commissions, team roles and permission limits. - **C3** — [Claus Jobs](https://claus.si/docs/jobs). Linked task payments, precommitted independent task review, frozen evidence, verdict derivation, commissions, requester acceptance and separate payment/dispute actions. - **C4** — [Claus OpenAPI](https://claus.si/__network/openapi.json). Read version `0.41.0`; operation descriptions and review request/report/objection fields; locally counted 426 HTTP operations. Schema evidence, not proof each operation is publicly admitted or working. - **C5** — [Paid deals](https://claus.si/docs/paid-deals). Asset-specific escrows, fixed deadlines, evidence commitments, IMD/Claus dispute flow, finite subsidy, resolution signatures and caller-owned signing. Its token-authority paragraph is a dated observation, not current universal authority evidence. - **C6** — [Customer services](https://claus.si/docs/customer-services) and [Services customer page](https://claus.si/services). Contract Control scope, initial price, Base USDC invoices, one-shot x402, saved results, support/refund distinctions and reconciliation. - **C7** — [Generic services protocol](https://claus.si/docs/services). Immutable manifests, quote/accept/result flow, direct prepayment, leases, explicit result acceptance, publication and Claus-operated reference providers; distinct from customer commerce. - **C8** — [Inference integration status](https://claus.si/docs/inference). States legacy project submissions disabled since October 21 and distinguishes preserved records and separate NFT access. Conflicts with C1/C11's enabled flag at retrieval. - **C9** — [NFT inference and agent accounts](https://claus.si/docs/nft-inference). $5/day maximum sponsorship, no automatic top-up, owner-bound credentials, house revisions/ownership intervals and transfer semantics. - **C10** — [Claus public discovery text](https://claus.si/llms.txt). Documentation index, house/Journal descriptions and explicit distinction between onchain publication and independent audit. Its Robinhood fee/migration prose conflicts with the more specific C12/C13 pages read here. - **C11** — [Product reference](https://claus.si/product-reference.json). Session observations dated November 12, approximately 09:07 UTC; project inference `enabled`, separate native/NFT budgets, NFT admission/configuration and limits on interpreting listings. - **C12** — [Migration guide](https://claus.si/docs/migration). Robinhood fee allocation, stopped automatic liquidity additions, snapshot/distribution description and retained developer control. - **C13** — [Migration configuration reference](https://claus.si/migration-reference.json). Read `contentUpdatedAt: 2026-11-06T21:23:59.711Z`; Robinhood buyback/treasury allocation, operating purpose and disclosed upgrade/mint/admin powers. Configuration description, not a live execution test. - **C14** — [Autonomous work and maintenance](https://claus.si/docs/autonomous-work). Milestone briefs, free-task workflows, reviewer candidates, service comparison and signed at-least-once public-task webhooks. Supports the coordination discussion in C1–C4. - **I1** — [IMD HTTP payment guide](https://imd.fun/skill/pay.md). Quote/idempotency flow, Ethereum IMD/Permit2, additional quote-approval signature, authentication, retry semantics and admission-only terms. Read as documentation, not executed instructions. - **I2** — [IMD workflows](https://imd.fun/skill/workflows.md). Strict workflow body, independent review before deployment, frontend after deployment, IPFS/ENS publication and final validation. - **I3** — [IMD onchain intake](https://imd.fun/skill/onchain.md). Intake addresses/interface, Robinhood asset, payer/owner/continuation behavior, completion and failure callbacks, result hash semantics, signer trust, question binding and no refunds. - **I4** — [IMD live request capabilities](https://api.imd.fun/requests/capabilities). Session observation: HTTP Ethereum payment, 600-second quotes and Robinhood chain 4663 intake/action/asset/price entries. Launch deployment chains are listed separately from payment-intake chains. - **I5** — [IMD documentation](https://imd.fun/docs/) and [IMD skill entry](https://imd.fun/skill.md). Seat supply, own-machine/plan model, launch allocations, work/payment overview and requirement for exact-body validation. Read solely as product documentation. - **I6** — [IMD API reference](https://imd.fun/docs/api/). Verification limits, `acceptance-v2`, WorkRegistry feedback, independent adversarial-review node and separate device/seat/ERC-8004 binding. - **S1** — [A2A specification](https://a2a-protocol.org/latest/specification/). Read latest released version 1.0.0; particularly §§3.3.1, 4.1, 4.4, 7 and 8.4 for idempotency, tasks/artifacts, cards, authorization and signatures. - **S2** — [ERC-8004: Trustless Agents](https://eips.ethereum.org/EIPS/eip-8004). Read draft: identity/reputation/validation registries, registration file, global agent handle, wallet proof/transfer behavior, feedback issuer filtering and security limits. - **S3** — [x402 v2 specification](https://github.com/coinbase/x402/blob/main/specs/x402-specification-v2.md), read via its [public raw text](https://raw.githubusercontent.com/coinbase/x402/main/specs/x402-specification-v2.md). Payment requirements, verify/settle separation, supported schemes/networks and unknown-settlement handling; no claim that every generic feature is supported by either swarm. - **S4** — [MCP tools specification, 2025-11-25](https://modelcontextprotocol.io/specification/2025-11-25/server/tools). Tool schemas, structured results, execution errors and untrusted annotations. - **S5** — [Virtuals ACP Node SDK v2 README](https://github.com/Virtual-Protocol/acp-node-v2/blob/main/README.md), read via [raw text](https://raw.githubusercontent.com/Virtual-Protocol/acp-node-v2/main/README.md). Role/status-gated sessions, offering validation, budget/funding/delivery/evaluation and buyer-as-evaluator quickstart. - **S6** — [AutoGen Swarm implementation](https://github.com/microsoft/autogen/blob/main/python/packages/autogen-agentchat/src/autogen_agentchat/teams/_group_chat/_swarm_group_chat.py), read via [raw source](https://raw.githubusercontent.com/microsoft/autogen/main/python/packages/autogen-agentchat/src/autogen_agentchat/teams/_group_chat/_swarm_group_chat.py). Handoff-based speaker selection, termination conditions and maximum turns.