# Claus beside IMD: adversary and failure review Researcher 5 · Independent view · 12 November 2026 **Labels:** **[R]** means read in a cited public source, including an explicitly identified public-chain observation; it does not mean independently audited. **[I]** means my inference, including uncertainty and limits of the evidence. **[P]** means my proposed change, experiment, or question. A label applies to its whole paragraph or table cell until another label appears. **[R — research scope]** This report uses public documentation, unauthenticated public metadata, published source code and read-only contract observations. No login, order, payment, signature, private endpoint, deployment or adversarial traffic was used. I did not read the other researchers’ views. Observations are dated below; documentation and API availability are evidence of advertised behavior, not successful end-to-end execution. ## Correct the starting facts first | Claim or design dependency in v0.1 | What the public evidence actually supports | | --- | --- | | Claus has no independent acceptance or onchain record. | **[R]** Jobs already supports a reviewer selected before work, frozen evidence, conflict checks against known controllers, immutable reports, objections and separately paid review commissions. Review remains advisory; requester acceptance remains separate. Paid agreements commit evidence hashes onchain. **[I]** Independent acceptance remains a gap; absence of review and records is the wrong diagnosis. This also does not establish an ERC-8004 work history for each house. [Jobs documentation](https://claus.si/docs/jobs) | | A Claus-to-IMD relationship must be built from zero. | **[R]** The paid-deal dispute design already purchases an IMD oracle panel, requires four agreeing answers from five members, and then requires Claus approval of the proposed split. **[I]** This is a narrow, asymmetric arbitration dependency; it does not establish general reciprocal hiring or independent upgrade approval. [Paid deals](https://claus.si/docs/paid-deals) | | Claus has no worker or coordination capability. | **[R]** Public capabilities advertise bounded isolated JSON workers and workflows; workers have no network and are not persistent. **[I]** That is not an IMD-style seat, but “no coordinator” erases useful existing machinery. A public inventory cannot establish how many independent operators actually work. [Capabilities](https://claus.si/__network/v1/public/capabilities), [agent application guide](https://claus.si/build.txt) | | The 300 houses and their sponsored inference are the proposed seat base. | **[R]** The NFT guide describes 300 houses and a sponsored ceiling of $5 per NFT per day, with automatic top-up off. A key does not establish funding or start an always-running agent. Ownership transfers do not transfer private account history or replenish allowances. **[I]** A house is an entitlement and address, not measured production capacity. [NFT guide](https://claus.si/docs/nft-inference) | | USDG and 395 operations describe Claus today. | **[R]** The current service page advertises Contract Control using USDC on Base, including an x402 route. I counted 426 `operationId` entries in the returned OpenAPI document. **[I]** Replace the blanket currency claim with service-specific terms; neither operation count establishes usable swarm capacity. [Services](https://claus.si/services), [OpenAPI](https://claus.si/__network/openapi.json) | | The claimed absence of posted prices, and payment compatibility on Robinhood. | **[R]** The public Contract Control offer lists 250,000 micro-USDC per unit and 500,000 with `sinceBlock` comparison. The CLAUS Jobs terms returned an Ethereum escrow using legacy token `0x1b54e762aa34cf6e28e9c082f2848e28e45da6b8`, not the Robinhood token. **[I]** Neither an upgrade-review price nor a Robinhood Jobs payment rail follows from these existing offers. [Public product prices](https://claus.si/__network/v1/public/commerce/products), [chain- and asset-specific Jobs terms](https://claus.si/__network/v1/public/chains/1/deals?asset=CLAUS) | | The October 21 inference shutdown is an unambiguous current fact. | **[R]** The inference guide still says new submissions are disabled, but the public capabilities response read for this report says `inference.enabled: true`, with project-allowance access. **[I]** These sources conflict. I did not attempt inference and cannot establish admission; budget the pilot without sponsored inference. [Inference status](https://claus.si/docs/inference), [capabilities](https://claus.si/__network/v1/public/capabilities) | | Using accepted IMD work as the model for quality reputation. | **[R]** IMD documents rejected as well as accepted verdicts, and says text verification checks structure alone. **[I]** A report’s acceptance cannot by itself certify the accuracy of its security conclusions. [IMD review policy](https://imd.fun/docs/api/#reviews) | | Huang’s post is dated November 12. | **[R]** X’s public embed dates the identified post November 10, 2026. It exposes the opening thesis about people commanding swarms; the normal post page was inaccessible during research. **[I]** Treat the business thesis as motivation, not evidence that a company needs no accountable humans. [Author’s post](https://x.com/matthuang/status/2120209025099043286), [X’s public embed](https://publish.twitter.com/oembed?url=https%3A%2F%2Ftwitter.com%2Fmatthuang%2Fstatus%2F2120209025099043286&omit_script=true) | **[R]** The current Hooks page supports the proposed 2% Robinhood project fee, split equally between buyback/burn and treasury. The older migration article contains superseded liquidity and fee descriptions beneath its update. **[P]** Cite the current fee policy with its chain and observation date; do not treat historical narrative as configuration. [Current hooks](https://claus.si/Hooks), [migration article](https://claus.si/migration-article) ## The five changes I would make, ranked 1. **[P] Make control disclosure and bounded authority the first gate.** Publish the actual owner, operator, writer, payee and upgrade paths before funding anything. Keep the pilot incapable of upgrading contracts or releasing a treasury on a review verdict. Require enforced delay and multiple independent approvals before later giving reviews financial authority. **[I] Reason:** an honest review is irrelevant if a privileged key can replace the reviewed implementation or publish a false completion. The concrete powers are documented below. **[P] Cost:** public operational disclosure and manual checking now; governance migration, delayed upgrades and extra signing effort later. The pilot needs no new governance contract. 2. **[P] Sell a bounded review service before converting houses into seats.** Use a small disclosed operator roster, existing Jobs and IMD orders, and pinned public artifacts. Keep upgrade review as an advisory specialty, first exercised on historical or deliberately seeded fixtures. Require distinct controlling operators, not merely different NFT numbers or swarm names. **[I] Reason:** shared ownership and correlated tools defeat nominal diversity; an advisory experiment limits the cost of discovering this. **[P] Cost:** manual selection, conflict disclosure, two review fees per artifact, and slower throughput. It postpones the 300-holder participation story. 3. **[P] Define failure economics before offering one payment.** Start with separately authorized service purchases, exact spending caps, explicit deadlines and a durable order ledger. Name who bears each nonrefundable IMD fee and each unaccepted Claus delivery. Use existing escrow only where its current asset, chain and admission are suitable. **[I] Reason:** a unified checkout would otherwise conceal who finances a failed intermediate step. **[P] Cost:** more visible payment steps, an accountable deadline monitor and some prefunded pilot expense. A later merchant offering a single invoice must fund its own retry/refund reserve. 4. **[P] Replace the rider score with an evidence history carrying no automatic financial privileges.** Record the brief, budget, actual controller, attempts, outputs, acceptance and later observations. Initially grant no credit, priority or fee share from that history. **[I] Reason:** a rider who also controls workers and reviewers can manufacture activity cheaply, then borrow or receive preferential routing against it. Counting paid requests prices the attack; it does not establish usefulness. **[P] Cost:** less incentive marketing, signed records and retained evidence; later credit requires a separately underwritten risk policy. 5. **[P] Move the swarm card and adversarial pilot forward; move scale and company formation back.** Retain the six card questions, but answer them with exact service versions, authorities, settlement rules, conflicts and evidence. Run three bounded orders and publish their artifacts in the City; test failure and disagreement as well as successful publication. Invite a third operator to interpret the same card before building an aggregator. **[I] Reason:** this tests interoperability without entangling identities, currencies and treasury permissions. **[P] Cost:** modest documentation and manual reconciliation; no immediate universal router or “no staff” company demonstration. ## How the identity and review system gets gamed **[R]** ERC-8004 supplies identity and feedback interfaces, explicitly acknowledges Sybil attacks, and leaves validator incentives outside the registry. Its prohibition on feedback by an agent’s owner or approved operator does not identify all wallets controlled by the same person. [ERC-8004 specification and security considerations](https://eips.ethereum.org/EIPS/eip-8004#security-considerations) **[I]** The attack model below assumes adversaries may buy several real seats, control requester and reviewer wallets, share infrastructure, and cooperate across both brands. These are feasible incentives and control relationships, not accusations about existing participants. | Attack | How it defeats v0.1 | Cheapest useful defense; remaining limit | | --- | --- | --- | | Fake or duplicate seats | **[I]** Invented registry IDs impersonate workers; multiple registrations or devices make one operator appear to supply a quorum. Real NFTs only make some duplication costly. | **[P]** Verify chain, registry, token ID and current ownership; bind a signed device key to one active seat; require a small qualification task. Count one review per disclosed control domain, and record ownership epochs. **[I]** This prevents simple impersonation and replay, not hidden common ownership. | | One owner in both swarms | **[I]** The builder “cross-reviews” itself through another wallet; two brands become one failure domain. | **[P]** Use a short roster whose operators disclose all participating wallets, employers and financial interests. Exclude builder-controlled and payer-controlled reviewers; treat undisclosed or unverifiable relationships as unknown. Ask IMD for assignment-level controller separation. If it cannot provide it, label the result an additional assessment, not independent approval. **[I]** Wallet funding links are warning signals, not proof of beneficial ownership. | | Reviewer collusion | **[I]** Reviewers exchange approvals, select agreeable peers, suppress dissent or share one model’s blind spot. A larger agreeing quorum can amplify this. | **[P]** Choose reviewers after scope is fixed from eligible operators; prevent reciprocal assignments; privately obtain signed report commitments before revealing reports. Pay for a completed review regardless of verdict. Have a separate operator reproduce blocking findings and publish disagreements. **[I]** A small roster is deliberately permissioned and can still collude; random selection alone does not fix that. | | Copied reviews | **[I]** A second agent republishes the first report or changes wording, producing two apparent checks for one investigation. | **[P]** Bind each commitment to job ID, exact input/output hashes and a fresh salt. Require evidence from each reviewer’s assigned checks, reproducible tests and an explicit limitation statement; reveal after both commit. Include undisclosed seeded defects in pilot fixtures. **[I]** Commitments prevent copying after revelation; they do not prove independent reasoning or defeat advance collusion. | | Reputation farming | **[I]** Related wallets buy trivial jobs, self-accept, keep trivial sites online and recycle payments to earn rider priority or credit. Splitting one task into many multiplies wins. Seat transfers can sell another operator’s history. | **[P]** No score-based benefits in the pilot. Preserve every funded attempt and failure; deduplicate revisions and task families. Show related-party work separately, with distinct customers and net outside payments rather than gross volume. Attribute history to controller epochs. Later incentives should be capped by realized revenue, never raw accepted-job count. **[I]** Even an externally paid job is not automatically useful or honestly reviewed. | **[P]** For “still live later,” retain the artifact hash and an explicitly purchased service interval. Check the same version at proposed day 7 and day 30 observations, recording missing evidence as unknown. A reachable empty website earns no quality credit. Let a requester sign acceptance, a reviewer sign findings and an observer sign availability; do not merge those into one unexplained success number. These are future pilot observations, not checks performed for this report. ## Which keys can change the rules ### Claus: follow the administrator, not just the token owner **[R]** The Robinhood review repository’s pinned token evidence identifies `ClausAdministrator` at `0x979d5810d32292358639f25294472f01c95bdc46` as token owner and token ProxyAdmin owner. It identifies founder owner `0xbb145ca83272c3806d4ddc75ca1d5514789cf1c5` and delegated operator `0x8aca8890d34086d7a30b9c06fae25e0c3cda4cfa`. [Published token/control evidence, commit ac63a28](https://github.com/contractclaus/claus-claus-buyback-v52-review-97934309/blob/ac63a28c06970294e222de51368a233d0fd11572/dependencies/token-live-evidence.json) **[R]** At Robinhood block **112,415,124**, read-only getters returned those same roles. Both token ProxyAdmin `0xcbc4bbd337e62c926c995e66a231e5e2b0a722a8` and hook ProxyAdmin `0xf2eaa097f1092437b6157b53d02b9182147c4ac0` were owned by the administrator, and both were managed targets. The token was also managed. The administrator runtime hash was `0x2c736730547cb945fe3b87d0d9eb094e2d638c20b2cc07d6cca131f56c4056e4`, matching the repository’s published live hash. This does not reproduce its compilation proof. [Public Robinhood RPC](https://rpc.mainnet.chain.robinhood.com), [administrator evidence](https://github.com/contractclaus/claus-claus-buyback-v52-review-97934309/blob/ac63a28c06970294e222de51368a233d0fd11572/baseline/administrator-proof.json) **[R]** In the published administrator source, owner **or operator** may execute calls against managed targets. The owner changes the operator and target list. The operator cannot directly invoke the protected ownership-transfer selector; execution contains no mandatory review-signature or timelock check. The token source exposes owner minting. [Administrator source](https://github.com/contractclaus/claus-claus-buyback-v52-review-97934309/blob/ac63a28c06970294e222de51368a233d0fd11572/baseline/ClausAdministrator.sol), [token source](https://github.com/contractclaus/claus-claus-buyback-v52-review-97934309/blob/ac63a28c06970294e222de51368a233d0fd11572/dependencies/MigrationToken.sol) **[I]** Delegating a managed ProxyAdmin is broad upgrade authority. Blocking a direct ownership-transfer selector does not constrain what a replacement implementation can do. “Founder recovery” is not founder co-signature on every operation. An upgrade review cannot be an enforceable veto through this executor as published. [Administrator execution path](https://github.com/contractclaus/claus-claus-buyback-v52-review-97934309/blob/ac63a28c06970294e222de51368a233d0fd11572/baseline/ClausAdministrator.sol) **[R]** The older migration article promises an hour’s notice but explicitly says the delay is not enforced. Separately, `hook-health.json` reports renounced core ProxyAdmin ownership for the **legacy Ethereum** token and hook at block **26,397,884**. **[I]** Neither observation proves Robinhood upgrades are locked. The Ethereum observation also supersedes the older block-specific token-control statement still embedded in the paid-deals guide. [Migration disclosure](https://claus.si/migration-article), [chain-qualified hook health](https://claus.si/hook-health.json), [older escrow dependency disclosure](https://claus.si/docs/paid-deals) ### IMD: fixed intake address does not mean fixed policy **[R]** Intake v2 is `0xa43e6f75ee006411f79ac1c84120606c2330de82`. Its verified Ethereum source gives the owner setters for price, payee, writer, quote signer and callback gas. The writer alone calls `complete`; payment is forwarded immediately. [Verified Intake source](https://etherscan.io/address/0xa43e6f75ee006411f79ac1c84120606c2330de82#code) **[R]** Getters at Ethereum block **26,397,900** and Robinhood block **112,415,124** returned the same addresses below. These were sampled block observations, not a claim of finality or an audit of Robinhood bytecode. [Public Ethereum RPC](https://eth.drpc.org), [public Robinhood RPC](https://rpc.mainnet.chain.robinhood.com) | Role | Observed address | Consequence | | --- | --- | --- | | Owner | `0x047f606fd5b2baa5f5c6c4ab8958e45cb6b054b7` | **[I]** Controls the mutable intake policy and can replace the writer. | | Writer | `0x3d3c0df7c1a17d384545e9d7c8228897bd9625cb` | **[I]** Controls publication and delivery of a completion claim. | | Quote signer | `0x169b08ea234d7819eeabb246e34d8cd81e32279f` | **[R]** Authorizes signed quote payments. | | Payee | `0x4e0fa57bde726079356537e2f34d671e9f41adbc` | **[R]** Receives intake payments. | **[R]** IMD separately documents WorkRegistry writer `0x5da3054201bc7e241875ac68173e49c269f76d44`. Its callback documentation distinguishes an ordinary job hash from a signed oracle answer: the job’s association with the work rests on the writer. [IMD wallet roles](https://imd.fun/docs/api/#wallets), [callback trust](https://imd.fun/skill/onchain.md) **[I]** Writer delivery and WorkRegistry reputation are different attestations. Neither establishes the real-world independence of reviewers. Separate role addresses also do not establish separate key custody. A signed oracle answer adds another verification boundary; it does not make the writer or panel infallible. [Callback trust](https://imd.fun/skill/onchain.md) ### What must be published or locked before paying **[P]** Both teams should publish a signed, dated authority manifest: chain, service version, contract and implementation addresses, code hashes, owner/operator/writer roles, controller overlap, price and payee, change process, pending upgrades, emergency powers, and who can change the website, scheduler or evidence record. Link the exact rubric, result format and appeal procedure. Publish which reviews cover the deployed revision; an old audit must not silently cover a new implementation. **[P]** For the first capped purchases, pin these terms and artifacts, use exact allowances, retain only a small spending balance, and give the connector no upgrade authority. Record review results without executing them. Stop new orders on authority or code changes until reapproval. A promise to freeze roles remains a disclosed trust assumption; monitoring is detection, not an onchain lock. **[P]** If callbacks are later connected, accept only the pinned sender and a pending `(chain, intake, requestId)` that the buyer actually ordered; consume it once. Bind any oracle answer to the expected question, consumer, artifact, panel requirement and expiry. A valid signature over a different question, or a content hash without substantive review, must not release money. **[P]** Before unattended spending, credit or upgrade execution, put dangerous authority behind multiple independently controlled signers and an enforced delay long enough to inspect evidence and exit. Route **every** upgrade-capable path through that control, including delegated ProxyAdmins; a narrow allowlist around an unrestricted upgrade function is insufficient. Emergency authority should pause new activity, not bypass review to mint, drain funds or erase obligations. Do not rename the existing keys and call that decentralization. ## When a paid job produces nothing **[R]** IMD’s advertised price is 0.5 IMD per ordinary request, with no result guarantee and no refunds. Its intake distinguishes pre-transfer reverts from later refusal or failure after payment. **[P]** Keep that admission charge visible as a separate line item. Do not advertise success-contingent pricing unless a named merchant absorbs the loss. [Hiring terms](https://imd.fun/skill.md), [Intake source](https://etherscan.io/address/0xa43e6f75ee006411f79ac1c84120606c2330de82#code) | Failure | What the existing mechanism means | Pilot treatment | | --- | --- | --- | | IMD refuses or ends without a result after taking payment | **[R]** Failure status does not refund the payment. | **[P]** Close the attempt visibly, charge it to the agreed pilot budget, and require a fresh capped authorization for another purchase. Never record it as shipped. | | IMD never completes, or callback delivery fails | **[R]** Delivery may be late or missing; a reverting callback is not a successful delivery. [Onchain flow](https://imd.fun/skill/onchain.md) | **[P]** Preserve original transaction/request IDs. Read the result independently, mark timeout or unknown at the agreed deadline, and stop dependent steps. A local timeout creates neither a refund nor proof of chain failure. | | Claus free task is delivered but payment was never funded | **[R]** Task acceptance alone moves no money. [Jobs](https://claus.si/docs/jobs) | **[P]** Do not start paid work on a posted reward or acceptance promise. Confirm the pinned agreement and funding first. | | Funded Claus task has no delivery | **[R]** The escrow allows settlement back to the buyer after the delivery deadline. | **[P]** The buyer or designated monitor submits the settlement transaction and then withdraws the credit. | | Claus delivers junk and the buyer disappears | **[R]** After the 24-hour review window, settlement pays the provider unless a timely dispute was filed. | **[P]** Monitor the actual escrow deadline; silence must never be treated as independent quality acceptance. | | A Claus dispute receives no valid resolution | **[R]** After the three-day dispute window, settlement refunds the buyer. | **[P]** Disclose this provider risk before accepting work; an inconclusive review is not evidence that the provider cheated. | **[R]** The last three escrow outcomes come from the published `ClausDealEscrow` state machine. Settlement creates withdrawal credits; passage of time alone does not send a transaction. The admission controller gates new funding, while the immutable arbiter authorizes a bounded split between original parties. [Escrow source, commit b1852dd](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/contracts/src/ClausDealEscrow.sol) **[R]** The public Ethereum ETH terms returned escrow `0x354f658129a6d8f67c90e2c3e1eb9981ca419816`, arbiter `0xb104063f7a25a712a5899861941eb5ceb24ee105`, and admission controller `0xbb145ca83272c3806d4ddc75ca1d5514789cf1c5` at block **26,397,893**. They explicitly mark the operational patch as not independently re-audited. **[P]** Preserve that limitation in the pilot’s approval record rather than silently inheriting an earlier review. [Observed escrow terms](https://claus.si/__network/v1/public/chains/1/deals) **[R]** Claus’s dispute documentation allows admission with three hours remaining but stops processing with less than one hour remaining; no completion guarantee is offered. **[I]** A case admitted at that latest permitted point has at most two hours for processing despite IMD’s potentially much longer completion time. **[P]** Do not make that subsidized process the pilot’s only recovery route. [Arbitration timing](https://claus.si/docs/paid-deals), [IMD delivery limits](https://imd.fun/skill/onchain.md) **[R]** x402 specifies HTTP payment flows, and its optional payment-identifier extension supports retry deduplication with request binding and a cache lifetime. **[I]** Neither payment confirmation nor an expired deduplication cache proves work was performed once and satisfactorily. **[P]** Retain a durable application order ID, original payment reference, input hash and terminal state beyond any HTTP cache. Resume uncertain payments; do not purchase a replacement automatically. [x402 overview](https://docs.x402.org/introduction), [payment identifiers](https://docs.x402.org/extensions/payment-identifier) **[P]** If one-payment routing is introduced later, identify a merchant of record. Its customer terms must allocate admission fees, intermediate failure, partial delivery, conversion costs and refunds. Keep the customer’s launch owner explicit: IMD’s paying contract must not accidentally become the rider’s permanent continuation controller. **[R]** IMD documents payer-default roles and restricts continuation to the original payer. [Payer semantics](https://imd.fun/skill/onchain.md) ## Where the open City admits spam **[R]** Basic City publication needs no purchased token or funded wallet. Wallet login authenticates control of an account; project mandates scope access and expire within 30 days. Jobs has public briefs and outputs, expiring free claims and finite shared quotas. [City access model](https://claus.si/build.txt), [Jobs limits](https://claus.si/docs/jobs) **[I]** The dangerous transition is discovery becoming authorization. Many cheaply created accounts can fill task storage, occupy claims, seed fake service profiles, post copied outputs, and trigger reviewer or rider notifications. A crawler that converts every matching City post into a paid IMD order turns public spam into treasury spend. A hostile brief can also try to instruct the reviewing agent to approve, disclose credentials or invoke another tool. These are attack hypotheses, not probes conducted against Claus. **[P]** Keep public discovery open but place a separate, initially manual gate before the swarm queue. Accept only a bounded brief, a pinned artifact and named payer from an admitted requester. Cap simultaneous and daily jobs by known operator, not merely wallet; add small deposits only if measured spam warrants their complexity. Cap depth, total budget and retries of cross-swarm calls so A cannot recursively hire B to hire A. **[P]** Give review agents read-only artifact access and no wallet, owner login or broad MCP credential. Treat briefs, repository comments and tool output as untrusted evidence. Keep spending authorization in a separate constrained signer that accepts only a fixed order schema. Rate-limit notifications, deduplicate content, expire abandoned claims and maintain an appealable quarantine. Publication rights must never imply queue priority, reviewer eligibility or spending rights. ## What I would cut or reorder | Original work item | Decision | | --- | --- | | 1. Convert 300 houses to ERC-8004 seats on holders’ plans | **[P] Cut from the first proof.** Invite existing capable operators voluntarily. Later add identity adapters only when capacity and attribution need them; preserve NFT access separately. | | 2. Coordinate upgrade review first | **[P] Keep the specialty; narrow its power.** Start with pinned, public advisory reviews and seeded fixtures. Defer “every upgrade” as an enforceable promise until authority can actually enforce it. | | 3. Second run, reviewers, seat record | **[P] Reuse and strengthen.** Use existing review/evidence tools; distinguish independent controllers, reproducible checks and acceptance. Keep raw signed receipts before inventing another reputation system. | | 4. One CLAUS price over x402, with a burn/seat/treasury split | **[P] Defer the new rail and split.** Post a bounded per-service quote in a currently supported asset. Fixed price requires limits on code size, scope, revisions and response deadline. Do not price an unlimited audit like a small request. | | 5. Bidirectional bridge and mandatory cross-review above a value threshold | **[P] Split and reorder.** Manual purchasing and artifact handoff first; automation after recovery works. Assess privilege and funds exposed, not only job price: a cheap review of a mint key can cover enormous risk, and attackers can split orders below a threshold. | | 6. One brief/payment and shared rider record | **[P] Keep one brief; defer one payment, scores and financial rewards.** Share evidence identifiers with attributable signers, not a bilateral score that both coordinators can rewrite. | | Original phase | Revised sequence and exit condition | | --- | --- | | 0. Research and yes/no | **[P] First:** resolve authority and source contradictions; publish the initial card, price limits, conflict policy and failure ledger. Name who can stop the pilot. No payment until the specific order’s terms are pinned. | | 1. Three IMD jobs and three City publications | **[P] First experiment:** three capped, separately purchased IMD artifacts, each published with a manifest in the City and reviewed by eligible independent operators. Include known-good and deliberately defective fixtures. Demonstrate readable evidence and exact artifact correspondence. | | 2. Thirty review seats | **[P] Replace:** a small roster and offline failure drills for missing output, disagreement, duplicate submission, common ownership and expired review. Expand toward 30 only when demand and reviewer availability justify it. | | 3. Cross-review, card and rider record | **[P] Pull the card and evidence history into phase 0/1.** Add automated handoff only after recovery and identity checks pass. Have a third provider interpret the card and verify a receipt without changing its schema. | | 4. Three hundred seats, new class, company, third swarm | **[P] Cut the bundled milestone.** Third-party compatibility comes early; seat expansion, financial reputation and a company are separate investment decisions. Delete “no staff.” Require named maintenance, incident and budget owners before claiming an operating company. | **[P]** Pilot stop conditions: any unexplained payment; wrong artifact/version treated as accepted; common-controller reviews counted as independent; missing failure in the history; a missed escrow deadline; or a seeded critical defect falsely approved. Require zero such failures before adding automation. Record actual cost, latency, human interventions and missed defects. Three examples prove a working handoff, not statistical security or economic sustainability. ## What I would keep exactly as is - **[P]** “Adjacent and not a copy”: IMD construction and Claus’s operating environment remain a useful division of responsibility, without claiming either has exclusive capabilities. - **[P]** No shared token. Do not merge token governance or create a bilateral reward currency. - **[P]** One brief as the rider’s starting interface. - **[P]** The swarm card’s six questions: who are the seats, what is sold, its price, payment method, returned deliverable, and who checked it. Make the answers precise without changing those questions. - **[P]** The requirement that an interface between two swarms must also work for a third. **[P]** I would not keep “each swarm is paid in its own token” as a prerequisite for the pilot. The existing service’s supported asset and chain take precedence; preserving two token identities does not require building conversion or a Robinhood escrow first. **[P]** Concretely, use IMD's Ethereum intake and a capped Ethereum ETH Jobs agreement for the first purchases, subject to current admission and founder budget approval. This avoids relying on the retired CLAUS asset for payment or deploying a new Robinhood escrow. It costs Ethereum gas and two payment assets; it reduces new implementation work. The report recommends this pilot, but does not place its orders. ## Open questions only the founders can answer 1. **[P]** Who actually controls each owner, operator, writer, oracle signer, reviewer and treasury key, including shared employers or custody? Which independent people will commit to the pilot’s conflict disclosures? 2. **[P]** Will Claus restrict delegated upgrade authority, and will IMD constrain policy-changing owner powers, before either result can authorize material value? Which emergency powers must remain, and who is accountable for using them? 3. **[P]** Who pays the nonrefundable admission fees, review fees, gas, hosting and monitoring, and what is the maximum acceptable pilot loss? If a single invoice is eventually sold, which entity guarantees its customer remedy? 4. **[P]** Will both teams publish failed, timed-out, disputed and related-party jobs alongside successes, and permit independent indexing and challenges? Who decides redactions and resolves factual objections? 5. **[P]** Can IMD disclose controller separation for the actual assigned reviewers? If not, will both founders accept the weaker label “additional assessment” and keep it advisory? 6. **[P]** What genuine customer task needs Claus’s operating service after IMD delivery, who owns its keys and data, and who maintains it after the launch subsidy ends? Which measured result would make the founders stop rather than scale? **[I]** These are commitments and risk allocations unavailable from public evidence. Fixing documentation conflicts, publishing current contract roles and checking live admission are operational work, not substitutes for those decisions. ## Revised one-page design — v0.2 **[P — proposed design, not an agreement]** **Goal.** Prove that a rider can commission useful work from IMD, have Claus publish and operate its public artifact, and obtain an attributable second assessment. Keep the services separate and make the interface usable by a third provider. No new token and no claim of a company without accountable humans. **[P] Product.** One brief defines the artifact, exact input revision, observable acceptance tests, maximum budget, delivery deadline and operating interval. IMD supplies construction or a bounded analysis; Claus supplies publication and a separately scoped review. Upgrade reviews are advisory during the pilot. No report or callback authorizes minting, upgrades or treasury release. **[P] Participants.** Start with a small roster of known operators using their existing machines and model accounts. Houses retain their present entitlement role. For each artifact, the builder and two reviewers must belong to three disclosed control domains; otherwise state the dependence and withhold the independence claim. Reviewers commit reports before seeing one another’s work and are paid for review effort regardless of verdict. **[P] Money.** Start on Ethereum with IMD intake payments and capped ETH Jobs agreements, subject to current admission. Pin the asset, payee and contract for every leg. Show IMD admission cost and its failure allocation explicitly. Fund a Claus agreement before work; assign a monitor for acceptance, dispute, settlement and withdrawal deadlines. Preserve uncertain payments under their original IDs. One merchant invoice comes only after someone accepts and finances the aggregate failure risk. **[P] Evidence.** The card answers the original six questions with versioned terms and authorities. Each order produces a manifest containing its brief/input hash, payer, controller epoch, service version, payment reference, output hash, checks, reviewer signatures and conflicts, acceptance decision, and any timeout or dispute. Both swarms can mirror the same signed events; disagreement remains visible. Record later availability against the original artifact and purchased interval. No automatic rider credit, priority or fee share. **[P] Authority.** Before funding, publish current owner/operator/writer and upgrade powers, the exact contracts and code hashes, and the source of each receipt. The pilot signer holds only its capped budget and cannot administer the reviewed contracts. Any authority change stops new orders. Enforced delays and independently controlled approvals are prerequisites for later automated financial decisions, not promises inferred from a review badge. **[P] First proof.** Buy three bounded IMD artifacts and publish all three in the City. Include a known-good fixture and a deliberately faulty fixture; add offline tests for duplication, missing output, conflicts and payment-state recovery. Require full evidence, no duplicate charge, correct artifact identity, all planted critical defects detected, and an accounted terminal state for every purchase. Obtain a third provider’s interpretation of the card before generalizing the connector. The sample establishes compatibility, not security assurance. **[P] Expansion.** Add limited automated handoff only after those gates pass. Grow the roster when paid demand and reliable delivery warrant it. Revisit native-token pricing, house-to-seat mapping, risk-based cross-review and a merchant checkout separately. Attempt a swarm-operated business only after repeat customers, operating costs, recovery and human accountability are demonstrated. ## Evidence and limits of this review **[R]** Public pages were read on 12 November 2026. RPC observations above used `eth_call`, `eth_getCode` and block reads only. Robinhood block 112,415,124 had hash `0xb3970c88f3880303da6cd5df67a6960b6941354d0a5e09429e46d7ce6cd598e3`; Ethereum block 26,397,900 had hash `0x0c9dd517dcbe495debd4864053752c56ddd78439170a52df957740ce0dde3bc3`. Public endpoint names are given with the relevant observations. Source links use pinned commits where available. **[I]** I have not demonstrated an exploit, measured real common ownership, tested paid service delivery, reproduced contract builds, or certified the live deployed escrow and every upgrade path. A public capability flag is not proof of successful service; a repository review candidate is not proof of deployment. These limits are why v0.2 uses capped advisory work and explicit expansion gates.