# Claus beside IMD: make independence and failure visible before scaling Researcher 1 · adversary and failure lens · public-source reading on 12 November 2026. **[Proposal] Verdict: yes to a small procurement experiment; no to v0.1's seat conversion, economic rider reputation or automatic cross-swarm approval.** Buy one bounded review of a frozen change, publish the evidence and the unsuccessful outcomes, and test recovery before selling a unified product. A different swarm name is not an independent security boundary. ## Reading key and limits **[Read]** means a public source says this, or the explicitly identified read-only RPC returned it. It does not mean the source's operational claims were independently audited. **[Inference]** means my conclusion from the cited evidence. **[Proposal]** means my recommendation, including all pilot quantities, gates and costs. **[Unknown]** means not established by this research. Labels apply to the whole paragraph or table cell unless another label intervenes. **[Read—method]** I read the primary sources linked below, including the public IMD worker repository and explorer-hosted Solidity source. I used public HTTP GETs and only read-only JSON-RPC methods (`eth_blockNumber`, `eth_call`, `eth_getStorageAt`, `eth_getBlockByNumber`). I did not log in, create identities, order work, send transactions, invoke private endpoints or use other researchers' views. IMD's skill.md was read as documentation, not executed. This report itself records the research scope and limitations; no separate README is delivered because only this output path is authorized. **[Unknown]** This is not a contract audit, proof of service availability, beneficial-ownership investigation or purchase test. I did not establish actual utilization, profitability, independence of deployed reviewers or the full Claus hook/vault authority graph. The exact Matt Huang primary post and its claimed November 12 date were not authenticated: primary-domain searches did not locate it. Treat the rider thesis as a hypothesis supplied by v0.1, not verified evidence for a business model. Nothing missing has been fabricated. ## Correct the premises first | v0.1 premise | Public evidence and correction | Consequence | |---|---|---| | Claus lacks independent acceptance or an onchain work record. | **[Read]** Claus Jobs already documents a precommitted reviewer, frozen evidence, immutable attributed reports, objections and linked ETH/CLAUS escrow commitments. Reports are advisory; only the requester accepts the task. App acceptance and financial settlement are separate. [Jobs](https://claus.si/docs/jobs) | **[Inference]** “No independent acceptance” is fair only if it means no independent reviewer with final acceptance power. “No onchain record” is too broad. Existing escrow commitments are not a universal seat reputation registry. Reuse the review primitive before building another. | | There is no bridge yet. | **[Read]** Paid-deal disputes already describe one IMD oracle request with five panelists and four agreeing exactly, followed by Claus approval of the split. Admission is conditional and completion is not guaranteed. Independent project launches also document an accepted independent IMD judge. [Paid deals](https://claus.si/docs/paid-deals), [independent launches](https://claus.si/docs/independent-token-launches) | **[Inference]** A documented integration exists; successful live operation was not tested here. A Claus judgment routed through IMD cannot automatically count as an independent second swarm judgment of IMD's work. | | Claus services use USDG and there is no posted price. | **[Read]** The current customer-services guide describes Base USDC payments and an initial Contract Control offer of 0.25 USDC per address; current offers remain authoritative. It also documents a direct Base-USDC x402 route. [Customer services](https://claus.si/docs/customer-services) | **[Inference]** The blanket USDG/no-price description is stale. A specific coordinated upgrade-review SKU still needs its own scope and price. Do not extrapolate one product's currency to all services. | | Sponsored inference makes the houses working seats. | **[Read]** NFT sponsorship is conditional on a funded period, allows up to $5/NFT/day and has automatic top-up off; a key does not start a persistent agent. Project inference is a separate service, closed to new submissions since October 21. [NFT guide](https://claus.si/docs/nft-inference), [inference status](https://claus.si/docs/inference) | **[Inference]** Neither house ownership nor a nominal allowance proves a worker is online, capable or economically viable. Closing project inference did not close every NFT allowance. | | ERC-8004 supplies independence. | **[Read]** ERC-8004 is a draft registry standard; its security section explicitly acknowledges Sybil reputation inflation and recommends filtering by feedback origin. Payment is outside its scope. [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004) | **[Inference]** Registration makes identities attributable, not unique humans or independent control domains. | | “Accepted” means independently proven correct. | **[Read]** IMD documents verifier receipts and profile-dependent checks; text verification checks structure, not substantive truth. [IMD API, Reviews](https://imd.fun/docs/api/#reviews) | **[Inference]** A valid report bundle can still contain a copied or false security review. Record which checks ran. | | All Claus contract-control statements describe the same deployment. | **[Read]** The URC-5 report identifies the Ethereum token and renounced Ethereum proxy-admin ownership. The Robinhood migration material describes developer control. My Robinhood reads below returned a nonzero proxy-admin owner. [URC-5 conformance](https://claus.si/urc5-conformance.json), [migration guide](https://claus.si/docs/migration), [migration article](https://claus.si/migration-article) | **[Inference]** Chain and address must accompany every authority claim. The build guide's broad “upgrade unavailable” wording must not be applied to Robinhood. | **[Read]** The current migration guide describes the 2% project swap fee as half buybacks/burn and half CLAUS DEV treasury, rather than a guaranteed compute/hosting allocation. [Migration guide](https://claus.si/docs/migration) **[Inference]** A treasury allocation is not a service-level commitment. Its destination and spend policy belong in the commercial terms. ## Five changes, ranked ### 1. Publish and constrain the authority graph before the first payment **[Inference]** The dominant risk is that a correct review can be bypassed by an authorized change, or that an authenticated receipt overstates delivery. Adding seats does not reduce either risk. The authority evidence below distinguishes observed addresses from the powers described in source. **[Proposal]** Each team publishes a signed, versioned deployment and authority manifest: chain, contracts, implementations and bytecode hashes; every upgrade/mint/fee/pause/admin role; owner and pending owner; operational writer and signer; treasury recipient; multisig threshold or timelock; and who can replace any of them. Include hosting moderation and result-publication powers separately from wallet authority. Name unresolved paths explicitly. Each pilot order pins that manifest version and artifact hash. **[Proposal]** For the pilot, use capped, individually authorized payments and exact allowances, with no pooled customer custody and no callback that releases capital or upgrades contracts. Before offering automatic spending or execution, move change authority to a disclosed multisig plus an enforced delay long enough for the job deadline and exit window; 48 hours is my proposed minimum, not a current guarantee. A pause may stop new purchases but must preserve existing escrow exits. Publish emergency bypasses; do not advertise a delay that an alternate role can evade. A changed manifest stops new dispatch until reapproved. Existing IMD requests still depend on its current writer: monitoring cannot freeze that dependency retroactively. **[Proposal—cost]** Small initial publication and review effort; ongoing monitoring and signer coordination. Enforcing a delay or changing ownership adds governance transactions, code review and slower incident response. Do not require irreversible renunciation just to run a tiny, explicitly trusted experiment. Refusal to publish or bound authority means no automated money path; a manually approved, disclosed-risk purchase can remain the experiment. ### 2. Replace “other swarm” with independent control and independently produced evidence **[Read]** IMD's public worker README says one NFT authorizes one active device and independent reviewers use different wallets. Claus explicitly warns that different identities do not prove separate real-world operators. [IMD worker repository](https://github.com/Identity-md/worker), [Jobs](https://claus.si/docs/jobs) **[Proposal]** Start with three disclosed operator groups, not thirty nominal seats: producer, checker and accountable acceptance owner. Require a signed common-control/conflict declaration covering both swarms and the customer. Exclude known common controllers across production, review and arbitration. Independently select the checker from a small vetted roster after the input is frozen; retain the selection record and prohibit requester-selected substitutions after seeing results. Disclose shared model/provider and control-plane dependencies. A repeat run by the same owner is a repeat, not an independent vote. **[Proposal]** Use sealed reports: each checker commits a salted hash of its complete review before any peer report is revealed. Require findings bound to the same source commit, compiled artifact, chain and state snapshot, with reproductions and explicit uncovered areas. A replay must state its environment and test inputs; mere agreement is not a test. The customer pays a fixed review fee for a valid negative or inconclusive review as well as a positive one. A disputed high/critical reproduction blocks the pilot release pending a named tie-breaker; inconclusive evidence never becomes a pass. **[Proposal—cost]** Manual roster/conflict maintenance, occasional third review, and extra review latency. Sealed reports need only signed files and timestamps initially. This is cheaper than a new identity or slashing system, but concealed common ownership and correlated model errors remain possible. The pilot's low exposure is the final defence, not a claim of Sybil elimination. ### 3. Sell an explicit failure contract, not “one payment” **[Read]** IMD's skill says payment buys admission, not a guaranteed result, and states no refunds. [IMD skill](https://imd.fun/skill.md) Its onchain guide says funds are forwarded immediately, status 1 is refusal, status 2 is no result, and callbacks can be late or missing. [Onchain guide](https://imd.fun/skill/onchain.md) **[Proposal]** Replace the unified payment with an itemized order: upstream nonrefundable attempt fee, separately funded Claus work, review fee, delivery deadline, acceptance window, dispute route and maximum retry spend. The payer must affirm who loses the IMD attempt fee if it produces nothing. Initially the sponsor pays that fee directly and absorbs it; unused downstream money is never automatically forwarded. A future reseller that promises a refund must fund that promise from its own explicit reserve, not imply IMD refunds. **[Proposal]** Keep separate states for payment confirmed, admitted, delivered, checked, accepted, credited and withdrawn. Preserve refusal, no result, timeout and disputed outcomes. No output by the chosen deadline is “no delivery by deadline,” even if the upstream job remains pending. Late output gets a separate event and no second payment without authorization. An unknown transaction outcome requires reconciliation using the original request ID/hash, not an automatic retry. The failure matrix below governs the pilot. **[Proposal—cost]** A clear order sheet, deadline monitoring and modest sponsor loss budget; customer friction from separate approvals. A refund-backed unified experience costs working capital, foreign-token inventory and reconciliation engineering. Defer those costs until repeat demand exists. ### 4. Replace rider scores with an evidence ledger; remove rewards for counting jobs **[Inference]** If accepted job count creates priority, credit and fee share, a rider can order trivial work between owned seats, buy reciprocal acceptance, and collect durable economic privileges. Even paid volume is weak evidence when the money cycles between related parties. Public receipts do not establish independent demand. ERC-8004's own Sybil warning supports treating raw feedback totals cautiously. [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004) **[Proposal]** Keep a rider activity log but grant no credit, fee share or scheduling boost in the pilot. One logical work item gets one canonical ID across swarms; revisions and retries remain children. Record related-party status, objective/rubric hash, original payment reference, producer/checker/controller identities, result hash, check profile, acceptance authority, failures and dated later observations. Both swarms mirror the same signed events rather than independently inventing scores. Conflicting attestations remain visible; the ledger can say “disputed.” **[Proposal]** If privileges are introduced later, count only reviewed outcomes from eligible independent counterparties, cap influence per controller/counterparty pair and time window, retain a denominator of eligible attempts, and separate attempted spend from net external revenue. Do not award uptime because a URL still resolves: bind “still live” checks to a version, functionality test and observation time. Changed code starts a new review. Owner changes start a new control interval rather than selling old performance as the new operator's skill. **[Proposal—cost]** A small event schema and manual review of relationships; slower growth metrics and no early token incentive story. Automated ownership inference is optional and uncertain. Do not deploy a transferable reputation asset to solve this bookkeeping problem. ### 5. Prove one review-and-publication handoff using existing tools **[Proposal]** Reorder the programme around one frozen, non-production upgrade candidate and one independent checker. IMD can produce a bounded build or review; Claus can commission a second analysis and publish the static evidence. Reverse the purchase direction only after that works. Do not start by approving a live treasury, token or vault upgrade. Existing Claus review and escrow features should carry the task where available; use manually signed evidence files where a feature is unavailable. No generic coordinator or contract bridge is required for the first proof. **[Proposal]** The pass criterion is not “six accepted jobs.” First recover or reproduce a known defect in a historical/seeded test candidate, correctly leave unsupported claims inconclusive, reject a known controller conflict, retain a failed-delivery record without duplicate payment, and reconcile delivered bytes with both public receipts. Label seeded examples as experiments and give them no economic reputation. Then obtain one real independent customer's bounded purchase and observe whether that customer orders again. A pilot can pass technical tests and fail the demand test. **[Proposal—cost]** One upstream fee plus explicitly quoted checker/publication work, operator time and test fixtures. These are budget categories, not measured prices or promised unit economics. The main opportunity cost is delaying seats and token routing. Add automation only for a handoff that repeats often enough to justify it. ## How people game it, and the cheapest defence **[Inference]** The attack paths below follow from transferable identities, advisory acceptance and public task inputs documented in [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004), [Claus Jobs](https://claus.si/docs/jobs) and the [IMD worker repository](https://github.com/Identity-md/worker). They are threat scenarios, not accusations that these attacks have occurred. | Attack and failure | Cheapest defence [Proposal] | Cost and residual risk [Proposal/Inference] | |---|---|---| | Fake or duplicated seats turn one operator into a supposed quorum. Buying more houses increases apparent consensus. | Check ownership and active binding at assignment; keep a cross-swarm controller declaration; count one independent control group, not one NFT, toward the review requirement. Admit a small accountable roster first. | Manual checks exclude some honest anonymous participants. Fresh wallets and hidden control can evade them. Device uniqueness alone cannot prove operator uniqueness. | | One owner holds seats in both swarms and reviews their own output under two brands. | Exclude known producer/customer affiliates across both registries; disclose agent, wallet and controller separately. Treat common control as one domain. | Disclosure is falsifiable attribution, not proof of completeness. Unknown control is explicitly unknown; it cannot secure high-value automation. | | Reviewer cartel rubber-stamps each other's work or extorts fees by inventing findings. | Independent assignment, no repeated reciprocal pair by default, sealed reports, fixed fees regardless of verdict, reproducible blockers and a named appeal reviewer. | Extra latency and occasional tie-breaker fees; a small roster can still be captured. Never pay per finding or per positive vote. | | Copied reviews create apparent independent agreement. | Salted commit/reveal before peer visibility; require artifact-bound reproductions, logs and explicit coverage; spot-check a sample by rerunning a claimed test. | Hashes prove commitment, not original thought. Private coordination and shared model blind spots remain. Identical correct boilerplate is not alone evidence of fraud. | | Riders farm reputation through cheap self-dealing, split jobs, retries or circular payments. | No financial privileges initially; canonical job lineage, related-party exclusion, counterparty caps and visible failure denominator later. | Some genuine repeat customers get less score weight. Fee burn creates friction but does not establish honest demand, especially if privileges are worth more. | | Seat/rider transfer or new wallet launders a bad record; selective publication hides failures. | Snapshot control per event, retain prior control intervals, publish every admitted pilot ID before work and reconcile terminal or overdue states. | No global identity system prevents re-entry. Pilot allowlisting and a bounded loss budget contain rather than solve it. | ## Authority: who can change what? **[Read—RPC]** The following are historical public-RPC observations, not machine information. Calls targeted the Intake at `0xa43e6f75ee006411f79ac1c84120606c2330de82`. Ethereum block `26397882` (`0x192ccba`, timestamp `2026-11-12T09:12:11Z`) was read through [Ethereum RPC](https://eth.drpc.org); Robinhood block `112413115` (`0x6b349bb`, timestamp `2026-11-12T09:12:23Z`) through [Robinhood RPC](https://rpc.mainnet.chain.robinhood.com). Both returned the same values below. Block numbers were decoded programmatically and timestamps fetched with `eth_getBlockByNumber`. These were pinned head observations, not a claim of finalized state. | Read method | Returned address on both chains [Read—RPC] | |---|---| | `owner()` | `0x047f606fd5b2baa5f5c6c4ab8958e45cb6b054b7` | | `pendingOwner()` | Zero address; no pending owner returned | | `writer()` | `0x3d3c0df7c1a17d384545e9d7c8228897bd9625cb` | | `signer()` | `0x169b08ea234d7819eeabb246e34d8cd81e32279f` | | `payTo()` | `0x4e0fa57bde726079356537e2f34d671e9f41adbc` | **[Read—source]** The Ethereum-verified `Intake.sol` uses `Ownable2Step`. Owner-only setters change prices, payee, writer, quote signer and callback gas. `complete` checks the writer address. `_collect` immediately forwards payment. The source shown has no upgrade function or timelock. [Verified Ethereum Intake source](https://etherscan.io/address/0xa43e6f75ee006411f79ac1c84120606c2330de82#code) **[Inference]** Writer compromise can misreport a normal job result or withhold completion; owner compromise can replace that writer and redirect future payments. Quote signing is a separate role from oracle signing. A content hash fixes bytes, not their correctness. I verified source on Ethereum; I did not establish identical Robinhood runtime bytecode. The human custody and signing arrangements behind these addresses remain unknown. **[Read—RPC]** For the official Robinhood CLAUS token `0x163015276f3244bC2CcB78b4D69cc8A4069EfFAC`, at block `112413344` (`0x6b34aa0`, timestamp `2026-11-12T09:12:48Z`), the same [Robinhood RPC](https://rpc.mainnet.chain.robinhood.com) returned: | Read | Value [Read—RPC] | |---|---| | Token `owner()` | `0x979d5810d32292358639f25294472f01c95bdc46` | | ERC-1967 admin slot | `0xcbc4bbd337e62c926c995e66a231e5e2b0a722a8` | | That admin's `owner()` | `0x979d5810d32292358639f25294472f01c95bdc46` | | ERC-1967 implementation slot | `0xd68727ddd50e8ba07e4be20d37a5ecf3dd814bff` | **[Read—method]** The slots were computed as `keccak256("eip1967.proxy.admin") - 1` and `keccak256("eip1967.proxy.implementation") - 1`, then fetched with `eth_getStorageAt` at the pinned block, following [ERC-1967](https://eips.ethereum.org/EIPS/eip-1967). Address results were decoded from the final 20 bytes of each ABI/storage word. **[Read]** Claus's migration article describes a single founder wallet with upgrade, fee, liquidity and extra-mint powers and a one-hour announcement commitment without an enforced wait. [Migration article](https://claus.si/migration-article) **[Unknown]** These RPC observations support concern about retained control, but do not prove the full current role graph, exact callable upgrade path, or that the observed owner is held as the article describes. Exact current hook and vault administrators were not independently established. Require their verified source and fresh role evidence before funding them. **[Read]** The Ethereum URC-5 report says its token/hook proxy admins have renounced ownership, with an observation at block `26397878`; the paid-deals page still describes an earlier Ethereum token-upgrade authority observation at block `26193874`. [Conformance report](https://claus.si/urc5-conformance.json), [paid deals](https://claus.si/docs/paid-deals) **[Inference]** These are different chains and/or observation times, not a basis for declaring all Claus assets immutable. A zero standard owner is also insufficient to rule out custom control paths. **[Proposal]** Before money crosses, Claus publishes exact hook/vault roles, the arbitration signer and scope, moderation powers, and whether upgrade reviews are enforceable gates or advice. IMD publishes custody/rotation of owner, writer, quote signer, oracle signer and WorkRegistry writer, plus handling of outstanding jobs during rotation. Both publish rate limits, data retention, content-addressed receipts and an exit runbook. Pin executable upgrade calldata and implementation hashes in a review; any change invalidates approval. Do not let a compromised writer's “success” callback authorize an upgrade. ## Paid work that returns nothing | Event | Existing rule [Read] | Pilot treatment [Proposal] | |---|---|---| | IMD transaction reverts before payment completes | The Intake rejects unsupported actions/insufficient price; its source forwards funds inside the transaction. [Intake source](https://etherscan.io/address/0xa43e6f75ee006411f79ac1c84120606c2330de82#code) | Record transaction failure and gas expense separately from a paid attempt. Retry only after identifying the cause. | | IMD payment succeeds, then admission is refused or work has no result | No upstream refund; failure status does not restore payment. A missing callback is also possible. [Onchain guide](https://imd.fun/skill/onchain.md) | Sponsor bears the disclosed attempt fee. Stop downstream spend, record the outcome, and request a new authorization for any retry. Do not infer success from silence. | | Claus task has no linked funded agreement | Posting, claiming and application acceptance do not pay anyone. [Jobs](https://claus.si/docs/jobs) | Never label a quoted reward “secured.” Do not start paid work until funding and the worker's onchain claim are confirmed. | | Funded Claus deal misses delivery | The buyer can recover the deposit through expiry settlement. [Paid deals](https://claus.si/docs/paid-deals) | Monitor the deadline; exercise settlement and then withdrawal. This recovers that escrow, not an already spent IMD fee. | | Provider submits empty/junk content and buyer disappears | After delivery there is a 24-hour accept/dispute window; silence makes payment available to the provider. Timely disputes have a three-day resolution window; unresolved disputes fall back to buyer refund. [Paid deals](https://claus.si/docs/paid-deals) | Assign a funded monitor and backup signer, check meaningful deliverables promptly, and dispute onchain. A review objection does not stop the clock. Providers must understand the fallback refund risk too. | | Escrow says credited, or a callback failed | Credit is withdrawable accounting, not wallet receipt; IMD results remain readable after callback failure. [Jobs](https://claus.si/docs/jobs), [onchain guide](https://imd.fun/skill/onchain.md) | Report credit, withdrawal and confirmed wallet payment separately. Keep original identifiers and make settlement recoverable without the City UI. | **[Inference]** x402 solves payment authorization/settlement over HTTP, not independent acceptance, asynchronous escrow or compensation when an upstream agent fails. Its documented flow places payment before serving the resource. [x402 introduction](https://docs.x402.org/introduction) **[Proposal]** A future one-payment router must identify its merchant of record, conversion/slippage policy, retry budget, refund reserve and partial-delivery rules. It must not acquire those obligations by accident. ## Where City openness enters the job queue **[Read]** Claus allows basic project/static publication without a funded wallet or NFT. Its open Jobs board accepts existing owner identities and delegated participants, with one-hour free claims. [Agent guide](https://claus.si/build.txt), [Jobs](https://claus.si/docs/jobs) **[Inference]** Spam enters through task posting, identity churn, claim-and-abandon loops, malicious briefs/evidence, public discovery profiles and recursive “hire another swarm” requests. A wallet signature proves control of an address, not work quality. Board-wide quotas can be exhausted even if each account stays within its own quota. A legitimate-looking service card can also advertise an attacker endpoint. These are attack surfaces, not demonstrated incidents. **[Proposal]** Keep the City open but make paid swarm dispatch a separate gate. Only a named pilot buyer and eligible funded task may trigger it. Validate scope, payload size and immutable input hashes; apply per-controller concurrency and spend caps; allow at most one outstanding pilot claim per operator; cool down repeated abandoned claims. No board post or profile automatically receives treasury sponsorship. Treat all briefs, source comments, review text and external pages as untrusted inputs; use isolated review execution without signing authority. No callback may recursively buy work without a separately capped authorization. A card's endpoint needs domain/identity binding and version pinning. Keep free discovery available without exposing it to an unrestricted paid queue. ## What to cut, retain and reorder ### Disposition of the six pieces | Piece | Decision [Proposal] | |---|---| | 1. Convert all 300 houses into own-plan ERC-8004 seats | Cut as a dependency. Houses remain houses. Invite willing operators independently of property ownership; registration can link identities later. Measure cost and availability before promising inference-funded work. | | 2. Coordinate upgrade review first | Keep the narrow scope, but start with a frozen historical or non-production change and manual assignment. Do not build a new general coordinator. | | 3. Independent run, reviewers and seat record | Keep evidence-based review; change independence to controller-level separation. Separate repeat execution, expert judgment, requester acceptance and registry writing. Record check scope rather than a blanket safety badge. | | 4. One CLAUS/x402 price and burn/seat/treasury split | Cut the new mandatory token/payment adapter and split for the pilot. Use the existing route's supported asset and separately posted price. A swap-fee burn is unrelated to the cost of a review. Revisit a CLAUS SKU after measured costs and demand. | | 5. Bidirectional bridge and value threshold | Keep ordinary procurement; defer autonomous contract ordering and mandatory paid callbacks. Replace a single value threshold with risk classes: permission-changing upgrades deserve review even with little current TVL. No second-swarm badge without independence evidence. | | 6. Rider interface and shared reward record | Keep a common brief and portable evidence ID. Cut credit, fee share and priority, and defer one payment. Record outcomes without turning counts into collateral. | ### Disposition of the five phases | Original phase | Revised order and gate [Proposal] | |---|---| | 0: research and yes/no | First: publish authority/failure manifests and identify an actual customer. “Yes” authorizes a capped experiment only. | | 1: three jobs each way | Start with one bounded handoff and failure-recovery exercise. Three reciprocal purchases are not independent demand. Expand to the proposed six only if evidence and economics justify them; label related-party purchases. | | 2: thirty seats | Defer. First use three disclosed control groups. Increase capacity only when the queue, diversity and recovery results warrant it. | | 3: cross review, card and rider record | Move a minimal card and raw event log before the first order. Cross-review is the first bounded experiment, not a later badge. No rider score. | | 4: 300 seats, second class, company, third swarm | Cut the seat-count target, second seat class and “no staff” company milestone. Move third-swarm compatibility testing earlier, after one proven handoff, to expose bilateral assumptions cheaply. Third-swarm production admission remains separately gated. | ### What I would keep exactly as is **[Proposal]** Keep these v0.1 choices unchanged: adjacent rather than copied swarms; no shared token; one job type first; phase 0's explicit yes/no decision; and the card's six questions—who are your seats, what do you sell, what does it cost, how do I pay, what do I get back, who checked it. The separate v0.1 requirement that every service be paid in its swarm token is changed for the pilot: existing USDC/ETH routes remain usable without changing either network's token. Keep third-swarm compatibility as a requirement. Everything else named above is intentionally changed or deferred. **[Proposal]** Preserve the six card headings exactly, but make their answers operational: chain-qualified identities and controller disclosures; versioned scope/exclusions; quote currency and expiry; payment finality/refund terms; schema/hash/deadline; and checker identity, method, conflicts and coverage. Attach authority/failure manifests as linked extensions. A card is discoverability, not admission or an endorsement. ## Questions only the founders can settle **[Unknown—founder decisions]** Public records can establish addresses and code, but not custody agreements, acceptable losses or business commitments. The following require decisions or nonpublic organizational facts, not guesses: 1. Who ultimately controls each owner, operational signer and review operator across both networks? Which beneficial-control overlaps will they disclose, and who vouches for that declaration? 2. Will Claus make review a technical prerequisite for token/hook/vault upgrades, or merely publish advice? Who can bypass it, for which emergency, and with what enforceable delay? 3. Will IMD and Claus move critical authority to disclosed multisigs/timelocks? Who owns rotation, incident response and monitoring of outstanding orders during a change? 4. Who pays the nonrefundable IMD attempt fee on upstream refusal, no delivery or disagreement? If a single merchant promises a refund, whose reserve bears it, and what is the approved loss ceiling? 5. Which genuinely independent buyer needs this review enough to pay, and what exact deliverable/rubric will that buyer accept? What reviewer fee is sustainable without treasury sponsorship? 6. Will owners voluntarily run Claus workers on their own plans, under what provider terms and duty expectations? Is participation independent of NFT ownership, and what rights do current NFT holders retain? 7. Who has final acceptance and appeal authority, who watches the escrow windows, and who is accountable for a malicious but formally valid delivery? What is the offchain liability policy for a missed vulnerability? 8. Does “rider owns it” mean contracts, repository, domain, revenue or legal entity? Who handles outages, renewals, compliance and security incidents? Will the founders abandon the “no staff” promise until those responsibilities have named owners? ## Revised design — one page **[Proposal: v0.2, all statements in this section]** **Goal.** Demonstrate that separately controlled operators can buy, check and publish one useful artifact across IMD and Claus with bounded financial loss and attributable evidence. Do not promise an autonomous company yet. The product hypothesis is a common brief and clear delivery record; willingness to reorder is the demand test. **Division of labor.** IMD supplies a bounded build or analysis through its existing intake. Claus supplies a separately commissioned review and City publication where available. The customer remains acceptance owner. Reviewer selection follows control independence rather than swarm membership. No new token, NFT conversion or general coordinator is required. **First product.** Review one frozen non-production upgrade candidate: source commit, chain, deployed-state snapshot where relevant, intended calldata, implementation/artifact hashes, risk questions and observable acceptance rubric. Deliver findings with reproductions, coverage limits and a pass/fail/inconclusive result for each criterion. No automatic execution of an upgrade or release of investment capital. **Trust boundary.** Publish chain-specific authority manifests before buying work. Producer and checker disclose cross-swarm common control and cannot be known affiliates. Use an independently selected checker, sealed reports and a named tie-breaker. Residual concealed ownership and shared-model risk remain disclosed. Critical authority needs an enforceable delay before future automation is permitted. **Money and failure.** Sponsor pays the disclosed nonrefundable IMD attempt directly; Claus review/publication has a separate supported-currency quote and, if available, the existing fixed-term escrow. State who bears every failure and retry cost. Start paid work only after confirmed funding/claim. Monitor delivery and dispute deadlines; distinguish acceptance, credit and withdrawal. Reconcile original transactions before retrying. No pooled custody, automatic swaps or refund promise without a funded guarantor. **Record and card.** Before dispatch publish the six-question swarm card plus authority/failure extensions. Both sides reference one logical order ID and signed, content-bound events: attempted, paid, admitted, delivered, checked, accepted, failed, disputed and later observed. Record controller intervals, related-party status and test coverage. Raw rider history earns no credit, priority or fee share. Permit contradictory attestations without silently rewriting history. **Sequence.** (0) Agree the rubric, loss cap and authority disclosures. (1) Prove one handoff plus deliberately labelled defect, conflict and non-delivery/reconciliation exercises. (2) Buy a real independent-customer task and publish all outcomes. (3) Test a third provider against the same card/receipt format without bilateral exceptions. (4) Add seats, payment routing or another product only when repeated demand and measured operations justify them. **Go/no-go.** Proceed only if the checker finds the known defect or reports insufficient evidence honestly, rejects the known conflict, receipts match exact delivered artifacts, failures remain visible, no recovery duplicates payment, and an independent customer buys useful work. Stop expansion on hidden authority, unverifiable independence, missing evidence or unfunded compensation obligations. A successful pilot establishes a narrow service, not general security or a staff-free company. ## Evidence access and reproducibility notes **[Read—access]** Several Claus documentation URLs and IMD skill/onchain pages were reported “not accessible via this tool” by the web reader. Immediate `curl -sS -A 'Mozilla/5.0'` GET fallbacks retrieved the documents used here. Etherscan's web-reader attempt also failed; curl returned HTTP 200 and the Solidity source. These were client-access problems, not evidence that either service was down. **[Read—limit]** The Robinhood Blockscout smart-contract API for the Intake (`https://robinhoodchain.blockscout.com/api/v2/smart-contracts/0xa43e6f75ee006411f79ac1c84120606c2330de82`) returned a Cloudflare challenge, including HTTP 403 on the browser-user-agent retry. No Robinhood verified source was obtained from it. Read-only RPC succeeded as recorded above. Exact Robinhood hook/vault powers, full implementation equivalence and live review availability remain unverified; none are represented as completed checks. **[Read—block anchors]** The block hashes returned for the three pinned RPC observations were, respectively: - Ethereum `26397882`: `0x2a4365d8c03e2f1dff7a423591b1a5bbf1eba1ffdb1b357014339ea0fdd42452`. - Robinhood `112413115`: `0x216577e2ab51c3c85ace6d2c68513b921e1bff7936ecdbe5f54e55b1d77e3935`. - Robinhood `112413344`: `0xc0f77465d3e91176461fa1390949143dd47efd9bc4f5fbba24cae70b57547eee`. **[Proposal—reproduction]** Re-read the linked public sources and repeat the named view calls at those blocks; recheck finalized state and verified bytecode before relying on the observations for a purchase. The report distinguishes documentation, direct observations and recommendations; it does not substitute a publisher's claims for independent assurance.