# Claus beside IMD: change the boundary before changing the seats Researcher 1 · Protocol and interface lens · Public-source desk research · 12 November 2026 **[Proposal] Approve a bounded interoperability pilot, not v0.1 as written.** Build a small contract for buying work between independent operators, adapt the review and payment machinery Claus already documents, and make every trust boundary visible. Do not make converting 300 houses into IMD-style workers a prerequisite. The first product should be an upgrade-review report with explicit scope; approval to deploy remains a separate decision. ## Reading convention and limits **[Read]** means a statement found in a cited public source, not a claim that I independently proved its implementation or operation. **[Inference]** means my conclusion from those sources. **[Proposal]** means my recommended design, including suggested costs, gates and questions. A label applies to its paragraph, list item or table cell; mixed paragraphs label their parts separately. **[Read—research record]** I retrieved public documentation, public capability and terms responses, and selected public repository files. I used no login, protected account, purchase, transaction, private endpoint or write endpoint. I did not read the other researchers' views. Dynamic observations below were returned on 12 November 2026, around 09:07 UTC. Public responses describe themselves; they are not independent uptime, deployment or security attestations. No live hiring experiment was performed. ## Correct the factual premises first | v0.1 premise | What the public sources actually support | Consequence | | --- | --- | --- | | Claus starts without coordination or independent acceptance. | **[Read]** The [Jobs guide](https://claus.si/docs/jobs) describes a reviewer and rubric committed before work, frozen evidence, authenticated reports, separate reviewer commissions and objections. The [contract-review guide](https://claus.si/docs/reviews-and-contracts) already describes source → build → independent review → Claus decision → execution → verification, including IMD audits. | **[Inference]** Existing coordination and advisory review should be reused. Enforceable independent acceptance is still an additional policy; advisory review alone does not fill that gap. | | Claus has no working seat. | **[Read]** [Build documentation](https://claus.si/build.txt) distinguishes NFT inference from an always-running agent. Separately, the [capability response](https://claus.si/__network/v1/public/capabilities) advertises isolated workers and durable runtimes. | **[Inference]** “No IMD-style holder-operated seat demonstrated” is defensible; “no execution machinery” is not. Advertised workers do not prove 300 active, independent operators. | | Cross-swarm review is wholly new. | **[Read]** The [paid-deals guide](https://claus.si/docs/paid-deals) describes an IMD oracle panel of five, four agreeing exactly on a payout, followed by a separate Claus decision and an escrow-specific resolution signature. | **[Inference]** Start by adapting and evaluating that existing integration. Do not count it as proof that IMD and Claus are independent for every assignment. | | Claus services uniformly use USDG and have no posted price. | **[Read]** The current [customer-service guide](https://claus.si/docs/customer-services) lists Contract Control at an initial 0.25 USDC/address, paid on Base, and a Base-USDC x402 route. The [service-order guide](https://claus.si/docs/working-together) calls its legacy priced-service path direct prepayment on Sepolia. | **[Inference]** There are several products and payment rails. Neither “all USDG” nor “no price” describes this surface accurately. A general CLAUS-priced upgrade-review product remains a proposal. | | Token migration implies all Claus payments moved to Robinhood Chain. | **[Read]** [Migration documentation](https://claus.si/docs/migration) identifies the current Robinhood CLAUS token and retired Ethereum source token. The observed [ETH deal terms](https://claus.si/__network/v1/public/chains/1/deals) remain on Ethereum; [CLAUS deal terms](https://claus.si/__network/v1/public/chains/1/deals?asset=CLAUS) also name Ethereum and the source token. | **[Inference]** Do not equate token location with service settlement support or reuse an escrow merely because its asset label says CLAUS. Migration compatibility needs an explicit decision. | | New project inference is definitely off today. | **[Read—conflict]** The [inference guide](https://claus.si/docs/inference) says new submissions were disabled on 21 October and `inference.enabled` is false. The retrieved [capabilities](https://claus.si/__network/v1/public/capabilities) instead return true, and the [product index](https://claus.si/product-reference.json) says enabled. | **[Inference—uncertain]** Current admission is unresolved by desk research. Do not test it by submitting a request or assume either document proves usable inference. | | Houses should become ERC-8004 identities in the same way as IMD seats. | **[Read]** IMD's [API reference](https://imd.fun/docs/api/#contracts) lists a seat NFT, a separate ERC-8004 adapter and external identity/reputation registries. Its [pairing documentation](https://imd.fun/docs/api/#pairing) binds a seat to an agent ID. | **[Inference]** Entitlement, agent identity and device authorization are separate objects already; Claus can map its existing objects rather than replacing them. | | The 2% trading fee supplies the proposed service split. | **[Read]** The [migration guide](https://claus.si/docs/migration) describes a Robinhood project swap fee split equally between buybacks/burns and CLAUS DEV treasury. | **[Inference]** This is a trading-fee policy, not evidence of worker compensation, a service margin or a review budget. | | v0.1 dates Huang's post November 12 and uses it to motivate a rider-owned, staffless company. | **[Read]** Huang's [original post](https://x.com/matthuang/status/2120209025099043286), retrieved through public HTTP, displays November 10, 2026. It emphasizes riders, evaluations and organizational design, including companies founded and populated by riders. | **[Inference]** Correct the date to November 10. The claim that the rider owns an entire company with no staff is v0.1's extrapolation, not an established protocol requirement. | **[Read]** IMD's [hiring skill](https://imd.fun/skill.md) confirms the flat 0.5 IMD admission price, no guaranteed outcome, holder-operated workers and HTTP/onchain purchase routes. Its [public docs](https://imd.fun/docs/) describe 2,000 seats and the 8% connected-seat/2% accepted-work launch allocations. **[Inference]** These are useful descriptions of IMD's product; none requires another swarm to reproduce its membership or economics. ## The five changes, ranked ### 1. Specify one purchase-and-receipt interface before any seat conversion **[Proposal—change]** Replace “publish a file answering six questions” as the interoperability specification with a versioned service profile, a binding order and portable signed receipts. Keep the six questions as the human introduction. Buyers should contract with a swarm service; its internal scheduler and membership remain its responsibility. **[Inference—reason]** Natural-language descriptions cannot determine whether retrying a lost response buys another job, which revision was reviewed, or whether completion means payment release. Claus's [service guide](https://claus.si/docs/working-together) explicitly separates delivery, acceptance and payment; [A2A's task model](https://a2a-protocol.org/latest/topics/life-of-a-task/) supplies asynchronous task identity and terminal states. Those are useful starting points for a commercial boundary. **[Proposal—minimum operations]** Require five operation families, mapped to existing endpoints rather than a new universal API: discover; quote; commit an order with its payment reference; read status/result; append a review or a buyer decision. Cancellation and disputes are explicit actions within the advertised lifecycle. Begin with polling; push delivery is optional. No operation requires access to the other swarm's scheduler or seats. **[Proposal—cost]** Each team must maintain one adapter, agree canonical schemas and keep recoverable order journals. This delays the attractive rider dashboard. In exchange it avoids coordinating 300 devices before the first useful purchase and lets a third swarm implement the same boundary without a bilateral special case. ### 2. Build acceptance on Claus's existing mechanisms, and separate it from review and release authority **[Read]** Claus's [OpenAPI](https://claus.si/__network/openapi.json) lists `reviewer_candidates`, task-review operations, contract reviews, service acceptance and deal arbitration. Reviewer matching can expose known shared ownership; its own description says distinct identities do not establish independent operators. The pinned [task-review implementation](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/src/apps/reviews.mjs) binds reports to task input/output, policy and agreement hashes and labels their authority advisory. **[Proposal—change]** Offer an `upgrade-review` service with an agreed rubric and named reviewers before the builder starts. Run reproducible checks separately from the builder, then obtain a reviewer report. For high-risk changes require a second disclosed operator from outside the builder's control, rather than assuming that a different swarm name is sufficient. The buyer or its explicitly authorized policy decides acceptance. Only the project's existing authorized signer can approve an upgrade. **[Inference—reason]** A repeated LLM run can reproduce the same omission. An onchain report hash establishes attribution and byte integrity, not correctness or reviewer independence. ERC-8004 itself [warns about Sybil reputation and unverifiable advertised capabilities](https://eips.ethereum.org/EIPS/eip-8004#security-considerations). An upgrade checker should be able to conclude “insufficient evidence” without losing its fee. **[Proposal—cost]** Pay for the independent review as a separate line item regardless of verdict. Budget additional latency, reviewer recruitment and dispute monitoring. Start with report-only reviews: enabling an automatic custody-changing upgrade requires another reviewed authorization design, not merely a passing report. ### 3. Use two bounded payment adapters and an explicit broker for “one payment” **[Read]** IMD's [HTTP payment guide](https://imd.fun/skill/pay.md) requires both a Permit2 payment signature and an EIP-712 quote approval; a stock x402 payer is insufficient. Its [live capabilities](https://api.imd.fun/requests/capabilities) advertise HTTP payment in Ethereum IMD and an intake on Robinhood Chain. Claus's current [customer-service API](https://claus.si/docs/customer-services) is a different service/payment product, not a generic paid hosting endpoint. **[Proposal—change]** Give each direction a durable adapter with its own spending wallet and finite budget. Let each service advertise the assets and networks it actually accepts. Keep “no shared token,” but remove the rule that every Claus product must immediately settle in CLAUS. Add a CLAUS rail only after the intended chain, token, transfer behavior and settlement implementation are explicit and reviewed. **[Inference—reason]** One customer payment cannot magically become two assets or bridge an asynchronous failure. An operator must hold inventory, execute conversion or accept credit risk. x402 supplies payment negotiation; its [network documentation](https://docs.x402.org/core-concepts/network-and-token-support) distinguishes protocol support from a functioning settlement path. It does not create cross-service atomicity. **[Proposal—cost]** Initially allow two disclosed subpayments within one brief. A later broker may offer one invoice, but must show conversion/slippage ceilings, gas, expiry, nonrefundable admission charges, its margin and who bears a failed second purchase. It must refund only money still under its control; it cannot promise to reverse IMD's spent admission payment. Inventory, reconciliation and a signer become real operating costs. ### 4. Map existing identities; keep riders' records separate from tradeable seat rights **[Read]** [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004) is currently marked Draft. Its agent identifier is the registry namespace/chain/address plus `agentId`; registration points to an agent description and service endpoints. It separates identity, feedback and validation from payment. ERC-721 ownership is transferable. Its wallet binding is cleared on transfer. **[Proposal—change]** Use ERC-8004 as an optional shared discovery and reputation anchor for a persistent agent or swarm gateway. Map existing Claus project/agent IDs and house entitlements to it. Do not register every short-lived run or automatically equate a house number with an agent ID. Preserve who operated an agent during each ownership/delegation interval. **[Proposal—riders]** A rider is initially a customer principal controlled by a wallet or organization account, with optional wallet-link proofs. Publish opt-in receipts for brief quality, paid orders, buyer decisions and later observations. If a rider operates an agent service, that service can receive its own agent registration. Do not mint a transferable human skill score or require customer registration to leave feedback. Use the same receipts at both swarms, not two independently edited rider histories. **[Inference—reason]** A purchased identity's historical output does not establish the new owner's competence. The ERC already [allows feedback from clients without requiring agent registration](https://eips.ethereum.org/EIPS/eip-8004#rationale). Credit, fee shares and priority would add gaming incentives before the record's meaning is established. **[Proposal—cost]** Identity mapping, key rotation history, consent and ownership-interval indexing take work. This sacrifices a simple NFT leaderboard, but avoids making customers buy membership or allowing reputation to silently change its human subject. ### 5. Replace population milestones and “no staff” with evidence gates for one bounded service **[Proposal—change]** Retain the six initial exchanges, but use their receipts and recovery behavior to decide whether to proceed. Next recruit a small, voluntary set of reviewers; thirty is a capacity ceiling to test, not evidence of success. Scale only after demand, reviewer independence, delivery latency and unit costs are measured. Admit a third swarm before declaring the interface stable, not after all 300 houses have been converted. **[Read]** The [Claus contract guide](https://claus.si/docs/reviews-and-contracts) limits its public reviewed contract profile and explicitly does not establish arbitrary token, vault or hook safety. [NFT sponsorship](https://claus.si/docs/nft-inference) is conditional on funded allowances, up to $5/NFT/day, with automatic top-up off; the key does not start a persistent worker. Anthropic's [multi-agent research account](https://www.anthropic.com/research/multi-agent-research-system) reports substantial token overhead and emphasizes clear delegation and outcome evaluations. **[Inference—reason]** Review effort depends on code, dependencies and value at risk, not house count. Moving inference costs to holders changes incentives and availability but is not an interop requirement. A company whose signer, hosting bill or review subsidy has no accountable operator has unresolved responsibilities. **[Proposal—cost]** Pay for measurement, incident response and responsible operator coverage. Price a bounded upgrade-diff product with explicit exclusions and additional quotes for larger work, rather than unlimited reviews at one arbitrary flat token price. The “staffless company” launch moves later; operational responsibility must be assigned before the claim can be tested. ## Answers to the interface questions ### The smallest hiring boundary **[Proposal]** Use the following five records. These are proposed wire requirements, not claims that either API already accepts this schema. Native order IDs, action versions and proof bytes remain unchanged inside each adapter. | Record | Minimum binding data | Who may assert it | | --- | --- | --- | | Service profile/card | Stable swarm/service identity; protocol and service version; operation/schema references; authorized keys; current availability/terms references; data policy; supported payment rails; review and settlement policy | The service operator, signed with a key linked to its identity | | Quote | Buyer and provider; service version; input/artifact hash; exact scope and output schema; permitted data recipients/publication; asset/network/amount/payee; delivery and review deadlines; acceptance, refund, cancellation and dispute rules; expiry | Provider; buyer commits to that exact quote | | Order receipt | Buyer-scoped idempotency key; quote hash; parent/root job IDs; native order/request IDs; payer/payment proof; admission outcome | The admitting service; payment finality comes from its stated payment rail | | Result receipt | Order and input hashes; output hash/URI/media type; immutable source/build references; completion time; retention/availability commitment; limitations | Delivering provider; delivery is its assertion | | Review/decision receipt | Exact result and policy hashes; reviewer identity and conflicts; evidence references; criterion outcomes; buyer's distinct decision; time and signature; supersession/revocation links | Each reviewer signs its report; buyer signs acceptance; an arbiter signs only within its separate authority | **[Proposal]** All new receipts carry schema version, issuer, intended counterparty, issuance time, signature key ID and a content digest with named hash algorithm. Use canonical JSON under [RFC 8785](https://www.rfc-editor.org/rfc/rfc8785), SHA-256 for the new document digest and a documented signature envelope. Represent large amounts as decimal strings. Keep a namespaced identifier for each swarm's native order; a root job links them without pretending their IDs are interchangeable. Preserve native IMD keccak commitments and Claus commitments in tagged proof fields rather than rehashing them as though they used the same format. **[Proposal]** Store work state and payment state independently: admission can fail after payment; delivery can succeed before acceptance; acceptance can exist before withdrawal. Include `unknown`, observed block/hash and finality where appropriate. A cancellation request is acknowledged separately from completed cancellation and any refund. Map each native deadline and failure code into a documented adapter outcome, retaining the original response. **[Read]** IMD's [onchain purchase guide](https://imd.fun/skill/onchain.md) says the request body is public onchain and reaches worker prompts. **[Proposal]** The pilot accepts explicitly public inputs only. Every quote names permitted onward recipients and publication rights; one brief does not grant unrestricted subdelegation. A hash commitment supplies integrity, not confidentiality. **[Read]** The [x402 payment-identifier extension](https://docs.x402.org/extensions/payment-identifier) describes request fingerprints and a cache TTL; after expiry the same identifier can be processed again. **[Proposal]** Therefore keep a durable application-level purchase ledger beyond the payment cache. Persist the intended order before signing. An unknown payment or broadcast holds the original operation for reconciliation; it never authorizes a replacement payment. Deliver event notifications at least once with deduplication, and retain polling as recovery. **[Proposal]** A third swarm conforms by implementing these records and mappings, not by adopting either token, house supply, model vendor or internal coordinator. Buyers can reject a service's trust/settlement policy without rejecting its transport implementation. ### Is the six-question card sufficient? **[Proposal]** It is the right length for a landing page and insufficient as a protocol. Add a seventh question: **“Who can authorize, change, revoke or dispute this service, and under which rules?”** That answer should include the signer/controller and upgrade authority, supported identity proof, key-rotation history, acceptance policy and arbitration/fallback policy. **[Proposal]** Make the original six answers precise: “seats” becomes an eligibility and delegation policy, not a mandatory public roster; “sell” references versioned input/output schemas and exclusions; “cost” references an expiring quote; “pay” names the chain, token, payee and rail; “get back” binds artifact hashes, availability and deadlines; “checked” identifies an actual report, its issuer, scope and conflict disclosure. A new service's card must say “no completed review yet” when applicable. **[Read]** [A2A discovery](https://a2a-protocol.org/latest/topics/agent-discovery/) already uses AgentCards for endpoints, skills, capabilities and authentication requirements; its [specification](https://a2a-protocol.org/latest/specification/#84-agent-card-signing) describes card signing. **[Proposal]** Publish an A2A-compatible card with references to this commercial profile where an A2A endpoint is genuinely implemented. Do not claim A2A support merely because a JSON file uses its vocabulary. Add a signed, timestamped capabilities document with expiry; a stale health observation is unknown, not a purchase guarantee. ### Exactly what ERC-8004 would register **[Proposal]** Register each persistent, accountable service identity chosen for the pilot: initially the Claus review gateway and IMD gateway if the teams want public anchors; then individual persistent reviewers when their identity needs to travel. The registration file names the service, describes its scope, links its real service/card endpoints, discloses supported trust mechanisms and links other registrations. Use an owner-controlled wallet and explicit scoped delegates; retain the existing house/project/seat identifier as a relationship, not a substituted token ID. **[Proposal]** Keep signed job and review receipts available offchain. Anchor issuer-attributed summaries or batches onchain as a later indexing layer, with enough pointers to recover the evidence. Tag scores by service/version and reporting party; preserve rejection, inconclusive review and subsequent failure. Do not describe an ERC-8004 validation response as payment authorization. For the rider, register only a rider-operated agent service if useful; keep the customer's record outside a mandatory agent NFT. **[Inference]** ERC-8004 is a useful shared reference, not a complete shared authorization system. A registry entry cannot establish that a session key may spend, approve an upgrade or speak for every member of a swarm. Its [security considerations](https://eips.ethereum.org/EIPS/eip-8004#security-considerations) explicitly leave capability truth and validator incentives outside the identity guarantee. ### Independent acceptance using Claus's present API **[Proposal]** Use this flow for the pilot; the evidence-only steps fit the existing Jobs review mechanism, while any automatic enforcement is additional work: 1. **[Proposal]** At task creation, commit the review policy, acceptance criteria and selected reviewer. Consult `reviewer_candidates` and accepted-work evidence, disclose controller overlap, and select deliberately. Do not treat a match ranking as an independence certificate. 2. **[Proposal]** Pin the old/new source commits, build settings, dependencies and relevant chain/block state. For an upgrade, include target chain/contract, current implementation and storage layout, proposed code/calldata, authority, invariants and recovery assumptions. The report approves only this scope; changes require a fresh review. 3. **[Proposal]** After delivery, freeze its exact bytes and evidence through `app_task_review_request`. A second operator reproduces the specified checks in a clean context; the reviewer reports per criterion through `app_task_review_submit`. Preserve `insufficient_evidence`, conflicts and limitations. Pay a commissioned review through its separate linked task. 4. **[Proposal]** Let `app_task_accept` record the requester's decision. Enforce any mandatory review gate in the buyer's own adapter or authorized signer, checking exact hashes and freshness before accepting payment or requesting deployment. Clearly mark this new gate as absent until implemented and tested. A review receipt alone never invokes an upgrade. 5. **[Proposal]** Monitor the original escrow deadlines; use an onchain dispute when money is disputed. A review objection is evidence, not the dispute transaction. Retain app acceptance, escrow credit and actual withdrawal as three separate observations. **[Read]** The pinned [escrow contract](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/contracts/src/ClausDealEscrow.sol) gives a delivered job a one-day review window and a dispute a three-day resolution window. On expiry, an undisputed delivered job pays the provider; an unresolved dispute refunds the buyer. Settlement creates credits withdrawn separately. **[Inference]** Silence cannot be labeled successful independent acceptance, and review delays can change financial outcomes. **[Read]** The observed [capability response](https://claus.si/__network/v1/public/capabilities) advertised arbitration enabled with seven remaining subsidized cases. The pinned [arbitration implementation](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/src/deals/arbitration.mjs) bounds the allocation at eight cases and requires time remaining for processing. **[Proposal]** Do not make this finite subsidy the guaranteed fallback for every job. Quote a named fallback and disclose what happens when no arbiter answer arrives. **[Proposal]** High-value independence needs a controller/conflict graph, reviewer selection insulated from the builder, reproducible artifacts and sampling against known defects. Use another swarm when it satisfies those conditions. If IMD built the work and supplies the arbitration panel, disclose that relationship and obtain a genuinely separate assessment where the risk requires one. Do not recurse forever into “review the reviewer”; state the final trusted decision maker. ### Calling across the boundary, and what keys must be trusted **[Proposal—IMD → Claus]** An authorized workflow extension invokes an external adapter after the IMD result is ready. The adapter pins the artifact manifest, obtains a current Claus service quote or funded Jobs agreement, checks its budget, pays with its own authorized signer, retains the original order and returns result/review receipts to the root workflow. For ordinary City publication, use a narrowly scoped project credential with the existing snapshot/release operations; a paid wrapper must specify what additional operating or retention promise the customer buys. The public evidence does not establish a ready-made paid hosting connector for IMD workflows. **[Read]** For a concrete existing paid service, [Contract Control](https://claus.si/docs/customer-services) exposes `GET /v1/x402/contract-control` through the public API origin, with an address parameter and Base-USDC x402 requirements. Its successful one-shot response contains the snapshot and receipt, without a saved customer invoice. **[Proposal]** Use that as the first paid-call adapter exercise and retain its response durably; treat paid hosting or a generic upgrade-review call as separate products to implement. **[Read]** [Claus's build guide](https://claus.si/build.txt) uses wallet-owner authentication and expiring project mandates; the [collaboration guide](https://claus.si/docs/working-together) says credentials must not be handed to another provider. Its project-owned service code has no general network/payment authority in the [OpenAPI](https://claus.si/__network/openapi.json). **[Inference]** A hosted app calling its local business logic cannot thereby spend a wallet or make an arbitrary paid external request. The bridge needs an explicitly authorized external payer or supported broker action. **[Proposal—Claus app → IMD]** A backend adapter, invoked under a finite user mandate, validates the exact request using IMD's documented preflight flow, then funds the Robinhood intake. For a contract-owned budget, use `requestWithFailure` with a narrowly configured receiver; pin intake, asset, action version and maximum price from capabilities. Persist the returned intake request ID and correlate it with the native job and completion receipt. Set launch `owner` and economic destinations deliberately, since the payer contract can otherwise become the default beneficiary. **[Proposal—Claus hook → IMD]** Keep paid work asynchronous. A swap hook emits or records a bounded work intent; an authorized keeper or owner-funded adapter processes it later. Do not put a model/network-dependent purchase or completion wait inside a trade. Deduplicate intents and cap purchases per trigger/window. The ordinary app and a hook use the same order semantics but different authorization entry points. **[Read]** IMD's [onchain guide](https://imd.fun/skill/onchain.md) says callbacks may be late or missing, get one bounded-gas attempt and can revert without blocking completion. It says job results are not oracle-signed: their hash fixes bytes while the writer asserts the association with the work. It separately describes oracle signature/question binding and unsigned failure callbacks. **[Proposal]** Store callback data, verify request ownership and sender, and perform consequential action later after the applicable evidence and review checks. Poll the original completion if a callback fails. | Boundary | Explicit trust and proposed restriction | | --- | --- | | Buyer mandate → local adapter | **[Proposal]** Trust the buyer's signer only for listed services/actions, assets, limits and expiry. Authenticate app requests separately from wallet spend. No caller-supplied arbitrary recipient or calldata bypass. | | IMD adapter → Claus | **[Proposal]** Trust Claus authentication, scoped credential enforcement, hosting availability and receipt assertions within the selected product. Give it neither the IMD seat/device key nor a treasury key. A dedicated payer can lose only its bounded budget. | | Claus adapter → IMD | **[Proposal]** Pin the intake and accepted result sender; trust the writer for ordinary completion attribution. For oracle decisions also verify the authorized oracle signer, bound question, consumer, quorum and expiry. Neither a callback nor a hash establishes code safety. | | Review → money/upgrade | **[Proposal]** Trust the selected reviewer only for its attributed report. Trust the named escrow arbiter only for its permitted payout split. Trust the existing project administrator for execution. Separate receipt-signing, spending, arbitration and upgrade keys. | | Key rotation or compromise | **[Proposal]** Record authorization intervals and revocations. Verify receipt keys as authorized at issuance; stop new orders on revoked keys. Explain that revocation does not reverse confirmed transfers or invalidate every already-issued financial signature. | **[Read]** The [x402 signed-offer/receipt extension](https://docs.x402.org/extensions/offer-receipt) supports portable signed artifacts and recommends a signing key distinct from the receiving wallet, including continuity through key rotation. **[Proposal]** Borrow that separation. Both swarms exchange public proofs and scoped requests; neither imports the other's private keys or assumes a shared root administrator. ## What to borrow from other systems | System and evidence | Borrow | Avoid or qualify | | --- | --- | --- | | **[Read]** [A2A](https://a2a-protocol.org/latest/specification/) defines cards, asynchronous tasks, artifacts, authentication declarations and multiple transport bindings. Its [task lifecycle](https://a2a-protocol.org/latest/topics/life-of-a-task/) treats terminal tasks as immutable and follow-ups as new tasks. | **[Proposal]** Reuse discovery vocabulary, task IDs, artifact references, explicit failures and related-task lineage. Keep the swarm behind the service boundary. | **[Proposal]** Do not equate A2A completion with commercial acceptance, or implement every transport before the purchase boundary works. Pin a tested protocol version. | | **[Read]** [Virtuals ACP's commerce description](https://whitepaper.virtuals.io/about-virtuals/commerce-layer) separates client, provider and evaluator around agreement and escrow. Its current [v2 SDK](https://github.com/Virtual-Protocol/acp-node-v2) uses event-driven job sessions and role-gated actions; its buyer example may select the buyer itself as evaluator. | **[Proposal]** Borrow explicit commercial roles, agreement-first execution and role-aware actions. | **[Inference]** An evaluator role is not evidence of independence. **[Proposal]** Avoid copying a registry, wallet stack or token dependency merely to inherit the role names; inspect actual evaluator selection. | | **[Read]** [MCP authorization](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization) concerns authorization for transport access, with resource/audience binding and scoped permissions. | **[Proposal]** Offer the adapter as a small set of MCP tools when useful, keeping credentials local and scoped. | **[Inference]** Tool access does not establish a paid-work agreement or independent review. **[Proposal]** Do not treat an MCP connection as wallet spending consent or silently forward its bearer token. | | **[Read]** [Anthropic's deployed research system](https://www.anthropic.com/research/multi-agent-research-system) uses a lead agent, bounded delegated work, parallel agents and outcome evaluations, with significant coordination and token costs. | **[Proposal]** Borrow precise briefs, output contracts, budgeted delegation and checks on outcomes. | **[Inference]** Coordination inside one provider does not solve independent operators' payment, identity or dispute problems. **[Proposal]** Leave those commercial boundaries explicit. | ## What I would cut - **[Proposal]** Cut mandatory conversion of all houses to holder-funded seats and mandatory public seat rosters. They are membership and compute-policy choices, not requirements for hiring a service. - **[Proposal]** Cut one undifferentiated CLAUS price for every request and any implied link from trading-fee burn to review affordability. Offer a bounded service and publish actual cost allocation before choosing margins or token subsidies. - **[Proposal]** Cut automatic high-value mutual-review loops. Require an independence policy and a named final authority; a network must be allowed to decline for lack of independent capacity. - **[Proposal]** Cut rider credit, priority and fee-share rewards from the pilot. Collect attributed records first, including failures and survival observations. A buyer's acceptance is not a neutral quality label. - **[Proposal]** Cut “the rider owns it, no staff” as a launch promise. State separately who owns code, contracts, upgrade rights, hosting accounts and ongoing bills. The protocol should expose accountable operators and recovery paths. - **[Proposal]** Cut a general blockchain bridge and compulsory dual-written reputation database. Use existing payment rails and portable receipts. Add cross-chain asset movement only for a separately justified purchase requirement. ## What I would keep exactly as is - **[Proposal—retain v0.1]** “Adjacent and not a copy.” Each swarm keeps its own implementation and membership rules. - **[Proposal—retain v0.1]** “No shared token.” Interoperability should not require a new joint asset. - **[Proposal—retain v0.1]** “One brief.” Keep a root request that can commission several explicit suborders. - **[Proposal—retain v0.1]** Upgrade review as the first coordinated job type. Its initial output is a scoped report and reproducible evidence. - **[Proposal—retain v0.1]** The phase-one counts: three IMD jobs ordered by Claus and three IMD builds published in the City. These are proposed future pilot transactions, not purchases made during this research. - **[Proposal—retain v0.1]** A link that works for a third swarm. Test that property with an independent adapter before freezing the interface. ## Open decisions only the founders can settle **[Proposal—questions]** These require commitments about ownership, budget or product intent; they cannot be settled from public endpoint lists. 1. Who is the seller accountable to the rider for the combined order, and who bears a spent IMD admission fee if the Claus leg fails? Is “one payment” worth operating an inventory-holding broker? 2. Which Claus product is being sold first: an advisory upgrade review, an approval policy, hosting, or continuous operation? What is the narrow scope and maximum exposure, and which existing keys retain final authority? 3. Which chain/asset is intended for new Claus review payments after migration? Will the existing Ethereum/source-token escrow be maintained, migrated or excluded? Who funds and reviews any replacement? 4. Does the NFT confer only current house/sponsorship rights, or also work eligibility? Will any machine/inference conversion require holder opt-in, and what happens to jobs and operator history when an NFT transfers? 5. Who qualifies as an independent operator across the two swarms? Are shared controllers, models, sponsoring wallets and previous work disclosed, and who handles the case where neither swarm can supply an independent reviewer? 6. Who supplies reviewer fees and arbitration capacity after the finite subsidy, and what service-level commitment can be made before a review deadline? Who monitors disputed jobs outside normal operating hours? 7. Which party owns each company's contracts, repository, hosting account, revenue rights and recovery mechanism? Is the rider an owner, customer, operator or some explicit combination? 8. Who may view or publish rider records, who signs factual observations, and can a customer opt out? If priority, credit or revenue shares arrive later, who assumes their loss and fraud risk? 9. Which interpretation of project-inference availability is authoritative, and which public deployment/terms document binds each service version? The current guide/capability disagreement must be resolved before relying on that capacity. ## Revised design v0.2 — one-page proposal **[Proposal—status and goal]** This is a proposal for separate teams to approve. Make Claus a persistent home, operator and reviewer of useful work beside IMD's build-and-launch network. A customer supplies one brief; named services undertake explicit suborders. Keep separate tokens and internal membership. Prove usefulness and recoverability before adding a joint company product. **[Proposal—first offer]** Sell a bounded upgrade-review report: exact old/new sources, build settings, dependencies, chain state and proposed change, against a committed rubric. Return findings, reproducible checks, conflicts, limitations and artifact hashes. A report cannot itself move funds or authorize an upgrade. The buyer's policy makes acceptance explicit; the project's existing signer retains execution authority. **[Proposal—boundary]** Publish a signed, versioned service card with schema/operation references, payment rails, authorized keys, live terms and review/dispute policy. Exchange five records: profile, quote, order, result, review/decision. Every order binds parties, input, price, deadlines and exact service version. Keep delivery, acceptance and payment status separate. Persist original operation and transaction identities through unknown outcomes. Status retrieval never creates a replacement purchase. **[Proposal—adapters and money]** Claus commissions IMD through an authorized backend or funded contract adapter; a hook records an asynchronous intent. IMD completion can trigger a scoped Claus publication or review adapter. Each payer has finite allowances and its own signer. Use existing supported assets and contracts. Start with transparent subpayments; offer one invoice only after naming a broker, conversion limits, gas, margin and loss/refund policy. Never promise to refund money already irreversibly spent downstream. **[Proposal—review and identity]** Reuse Claus task-review, evidence, commission and dispute mechanisms. Select reviewers before work; disclose common control and pay for review regardless of verdict. Treat cross-swarm review as one way to obtain independence, not proof of it. Use optional ERC-8004 identities for persistent services/reviewers, mapped to existing project and membership identifiers. Keep rider records opt-in and receipt-based; defer financial rewards. Preserve operator/ownership intervals and key history. **[Proposal—stages and gates]** Stage 0: agree product scope, seller, supported chain/asset, trusted authorities and failure rules. Stage 1: perform three Claus→IMD orders and publish three IMD builds in the City; retain linked receipts, verified payment observations and buyer decisions. Stage 2: pilot upgrade reviews with voluntary reviewers and seeded-defect evaluations. Require demonstrable no-duplicate-purchase recovery, stale-evidence rejection, timeout/dispute handling and clear accounting of costs and missed deadlines before scaling. Stage 3: have a third independent swarm conform using only the published profile and run the same recovery cases. Scale toward thirty and then 300 only when demand, budgets and independent capacity justify it. **[Proposal—company gate]** Launch a joint company only after assigning asset ownership, upgrade and spending authority, operating budgets, incident responsibility and an export/recovery path. Observe whether accepted work remains available and useful at stated later dates. Record unknown observations honestly. The deliverable of the pilot is a usable commercial boundary and measured service, not a seat count or a promise of unattended permanence. ## Evidence and checking notes **[Read—research record]** Claims above rely on the linked first-party network documentation, public APIs, specifications, repositories and Huang's original public post. Claus code links are pinned to public commit `b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f`; other live pages can change. I found no public evidence sufficient to certify independently operated Claus seats, new generic CLAUS-paid upgrade-review capacity, a production paid-hosting connector for IMD workflows, or staffless company ownership; those remain proposals or open questions. **[Read—local checking record]** The deliverable was checked locally for required sections, explicit claim labels, Markdown links, text encoding and the requested output path. A public HTTP check retrieved every cited URL successfully and checked the cited HTML fragments. HTTP success does not certify the content's truth. No independent reviewer, protocol implementation test, paid service execution, deployment audit or chain-state verification was performed.