# Claus beside IMD: how it breaks, and what to cut **View 3 of 5. Lens: adversary and failure.** This is a desk review of DESIGN v0.1 (ultrapara). I only read public pages, on 2026-11-12. I did not log in, sign anything, or call any write or private endpoint. The only reads were public GET pages and JSON listed in the sources. **How each statement is marked** - **[R:Sn]** means I read it in source *n*, listed at the end. - **[I]** means I inferred it from what I read. - **[P]** means it is my own proposal. - **[?]** means I could not verify it. --- ## 0. Facts to fix first v0.1 gets several facts wrong, or relies on things it cannot back up. These have to be fixed before any of the design can be judged. | # | v0.1 says | What the public sources say | Mark | |---|---|---|---| | F1 | The pool fee is 2%, half to buy and burn CLAUS and half to compute and hosting | The Claus sources disagree with each other. `migration-reference.json` (updated 2026-11-06) gives 1% buyback-and-burn plus 1% operating treasury on Robinhood Chain. The homepage says the same in prose. But `llms.txt` says the Robinhood 2% goes *all to liquidity*, with "no current treasury share". On Ethereum the official pool charges **2.3%**, and all of it goes as ETH to one fixed recipient (`0xBb14…F1C5`). | [R:S7][R:S8][R:S6] | | F2 | (implied) the split is a stable property that a posted price can rely on | The Robinhood token, hook and liquidity can all be upgraded by the owner. The owner is one wallet (`multisig: false`). There is **no onchain timelock**. Extra minting is allowed with **no cap**. The promise of "1 hour notice" is not enforced by the contracts. | [R:S7][R:S6] | | F3 | The 300 NFT houses can "become seats" | The house collection is on **Ethereum** (`chainId 1`, collection `0x8039…95d9`). Each NFT is backed by 50,000 CLAUS, and **the current owner can burn it to get those tokens back**. Selling one does not move the house's private projects or keys to the buyer. | [R:S14][R:S6][R:S12] | | F4 | Houses have sponsored inference of up to $5 a day | Confirmed: up to $5 per NFT per day. But the live profile is `claude-haiku-5.5`, with at most 1,024 output tokens, 6 requests a minute, one request at a time, and `supports300Independent: false`. | [R:S12][R:S13] | | F5 | New project inference was switched off on Oct 21 2026 | Confirmed for *project* inference. NFT-sponsored inference is a separate service, and it is admitting requests. | [R:S15][R:S13] | | F6 | Claus payment is "a separate step", implying no protection | Payment is a separate step, but it goes into an **immutable escrow**. Delivery must happen within 30 days, and `settleExpired` refunds the buyer if nothing is delivered. After delivery the buyer has 24 hours to accept or dispute, and silence pays the worker. A dispute gets 3 days with a named arbiter, and if that runs out the buyer is refunded. The escrow itself cannot be upgraded, but **the CLAUS token it holds can be**. | [R:S10][R:S9] | | F7 | IMD takes x402 with Permit2 or contract intake "on Ethereum and Robinhood Chain" | Confirmed for the intake (same address `0xa43e…de82` on both chains). Most of the live HTTP price list names `eip155:1`: 9 of the 11 network entries are Ethereum and only one is Robinhood Chain (4663). | [R:S1][R:S5] | | F8 | "Onchain record" (both sides) | IMD writes its reputation on **Ethereum only**, through the WorkRegistry writer into the ERC-8004 ReputationRegistry. The ERC-8004 registries are listed for Ethereum and Sepolia, not Robinhood Chain. Claus has no ERC-8004 record at all. Its receipts are input/output hashes in its app, and Journal records on Ethereum. | [R:S2][R:S9][R:S6] | | F9 | Riders' records "earn priority" | Today IMD routes work by tools, load, failures and turn. Its docs say **"past wins count for nothing"** and "nobody reviews their own work". Giving priority for reputation would be new, and it would create the main reason to farm reputation. | [R:S4] | | F10 | Matt Huang's post of Nov 12 2026 on swarm riders | **I could not find it.** A web search for "swarm riders" or "swarm native" with his name found nothing. Treat the thesis as the proposer's own claim, not a citable fact. | [?] | | F11 | "395 API operations plus MCP" | I did not verify the count. The OpenAPI schema and MCP clients exist. | [R:S6][?] | | F12 | ERC-8004 as a settled standard | It is a **DRAFT**. The ERC says outright that "Sybil attacks are possible" and that unfiltered summaries are "subject to Sybil/spam attacks". It also requires callers of `getSummary` to name which reviewers they trust (`clientAddresses`). | [R:S18] | --- ## 1. Answers to the lens questions ### 1a. Fake or duplicate seats **How it breaks** - IMD seats are scarce: there are 2,000 NFTs [R:S2]. But one person can run many of them. A public community repo runs "several NFTs on one server, one container each" [R:S4, which links S20]. So "different seat" does not mean "different operator" [I]. - On Claus, a house is a 50,000-CLAUS NFT that can be burned back into tokens [R:S6]. That makes a seat a tradeable, cash-redeemable claim on future fee share. Someone can buy one, farm the record, then sell or redeem it [I]. - Claus agents can also be republished across projects owned by the same person. Claus labels that link as "owned", which "is not independent collaboration or review" [R:S11]. **Cheapest defences** - [P] In any cross review, count *operators*, not seats. An operator is the wallet that holds the NFT plus any wallet that funded it within N days, and that link can be read from public chain data. One operator gets at most one vote per job. - [P] Use the *existing* scarcity as the stake. A seat that is transferred, or a Claus house that is redeemed, starts its record again for priority purposes. ERC-8004 feedback stays onchain, and readers filter by `clientAddresses` anyway [R:S18]. ### 1b. One owner holding seats in both swarms **How it breaks** v0.1's strongest anti-collusion rule is "work above a set value gets its second review from the other swarm". That rule fails quietly if the same person holds an IMD seat and a Claus house. They can build on one side and approve on the other [I]. Claus's own escrow docs say: "Different addresses do not prove independent operators" [R:S10]. **Cheapest defence** - [P] Make each seat publish a self-signed cross-holding statement, binding its IMD agentId to its Claus house tokenId. - [P] Exclude any declared or chain-linked pair from reviewing the other's work. - [P] Treat a link that was not declared but is found later as grounds to drop that reviewer's past reviews from the record. Do not try to slash money: nothing is staked [I]. This costs a published file per seat and a funding-graph check in the coordinator. It does not need a contract. ### 1c. Reviewer collusion **How it breaks** - In IMD's panels the network picks the reviewers, and a seat cannot review its own work [R:S4]. But many seats behind one operator can still end up on the same panel [I]. - On Claus the requester chooses the reviewer and commits to it in `reviewPolicy` before anyone claims the task [R:S11]. So a requester and reviewer who are in league can certify anything. **Cheapest defence** - [P] For cross reviews, the *other* swarm's coordinator picks the reviewer, not the requester. On IMD that is the control plane's existing panel choice [R:S4]. - [P] Pay the reviewer whatever they conclude. Claus already recommends this [R:S11]. - [P] Count only findings that come with a reproduction. IMD already blocks only on "high or critical with a reproduction" [R:S2]. A reproduction can be checked again, so a reviewer who invents or hides findings can be caught by a single honest rerun [I]. ### 1d. Copied reviews **How it breaks** A second reviewer can read the first review, which is public once delivered: Claus briefs, outputs and IDs are public [R:S9]. They then submit a paraphrase and earn credit without doing the work [I]. An "independent second run" (piece 3) is worthless if it can see the first run's output. **Cheapest defence** - [P] Commit and reveal. Each reviewer first submits the hash of their report. Reports are revealed only after all hashes are in. - [P] Both sides already have the pieces. Claus has evidence documents hashed with SHA-256 [R:S10], and IMD has a result hash on the intake callback [R:S3]. - [P] When findings are identical, credit the earliest commitment. ### 1e. Reputation farming **How it breaks** - The rider record and seat record in v0.1 count "jobs ordered, accepted, shipped" [R: v0.1]. On Claus, free tasks need no funded wallet and only the requester accepts [R:S9]. So one person can post, claim and accept their own tasks with two wallets, and pay nothing [I]. - On IMD each request costs 0.5 IMD [R:S1]. Farming is not free there, but it gets cheap if 0.5 IMD is worth little [I]. - An ERC-8004 summary that is not filtered by trusted reviewers is open to Sybil attack by design [R:S18]. **Cheapest defence** - [P] A rider or seat record counts only **paid** jobs. "Paid" means through the IMD intake, or through Claus escrow ending in `credited` [R:S9]. - [P] The payer must not be chain-linked to the worker. - [P] Weight each job by what was paid, and by whether the result is **still live after 90 days**. "Still live later" is already in v0.1; make it the main signal, because it is the expensive one to fake. - [P] Readers filter by a published list of trusted requesters, which is how ERC-8004 is meant to be used [R:S18]. - [P] Do **not** give routing priority for reputation until this filter has run for one phase. IMD's current "past wins count for nothing" rule [R:S4] is a strength, not a gap. ### 1f. Which keys can change the rules **Claus** - One owner wallet, `0xBb145cA83272c3806D4dDc75cA1D5514789cF1C5`, can upgrade the Robinhood token, hook and liquidity, move liquidity, and mint with no cap. There is no multisig and no timelock [R:S7]. - The Ethereum CLAUS sources contradict each other: - `llms.txt` says ProxyAdmin ownership was renounced and only metadata can be changed [R:S6]. - The paid-deals page says "CLAUS itself can" be upgraded, with ProxyAdmin `0xfb4b…7a57` owned by `0xBb14…F1C5`, observed at block 26193874 [R:S10]. - [I] The later renunciation may explain the difference, but a reader cannot tell without checking the chain. - The Jobs escrow cannot be upgraded, but its named arbiter is a trust assumption [R:S10]. - The NFT inference service's terms are versioned and controlled by Claus (`2026-10-23-v1`) [R:S12]. **IMD** - The Intake writer `0x3d3c…25cb` is "the only caller of the Intake's `complete`". IMD states that this key "cannot forge an answer but can withhold one, pair one with the wrong request … or send a false failure" [R:S2][R:S3]. - The Quote signer, the Intake payee `0x4e0f…bc` (all payments are forwarded there in the same call), the WorkRegistry writer `0x5da3…6d44`, and the Launch deployer are all single wallets [R:S2]. - Intake addresses are fixed. A new version is deployed at a new address and the old one keeps being served [R:S3]. - **I found no public statement of who owns the Intake contract, or whether it has an owner, pause or price setter** [?]. That has to be published. **What each side should publish or lock before money crosses** [P] - **Claus:** - Move the Robinhood owner to a multisig with a published signer set. - Put upgrades, mints and fee changes behind an onchain timelock of at least 72 hours. One hour is shorter than a review job; IMD results arrive "minutes to days" [R:S3]. - Put an onchain cap on extra minting. - Fix the F1 inconsistency, with a single machine-readable fee state. - **IMD:** - Publish the Intake's owner and admin functions (or state that it has none). - Publish how the writer and quote signer keys are held. - Publish an onchain-readable price, which `intakePrice` partly provides [R:S3]. - **Both:** - A swarm card field: "who can change the rules, and with how much notice". ### 1g. A paid job that returns nothing **IMD** - A refused or failed request is a spent payment, and there are no refunds [R:S3]. - A status-2 "no result" outcome can mean the panel disagreed, or the job was cancelled, blocked or superseded. It also costs the payment [R:S3]. - Failure callbacks are unsigned, and anyone can cause one by sending a bad body [R:S3]. - So a Claus contract that orders IMD work must never move value on a failure [R:S3]. It should treat 0.5 IMD as a sunk entry fee [I]. **Claus** Free tasks have no money attached. For paid tasks the escrow *does* refund an undelivered deadline (`settleExpired`) or a dispute nobody resolved [R:S10]. But there are two traps: - **Silence pays the worker after 24 hours.** A buyer who does not watch the deadline pays for nothing [R:S10]. - **Moving to escrow is optional** and happens only after a claim [R:S9]. **Cross-swarm** The rider pays "one payment" in v0.1. If that payment is split into a non-refundable IMD part and an escrowed Claus part, the rider carries the IMD loss and the Claus deadline risk without knowing it [I]. **Proposal** [P] - Run every quote through IMD's free check and quote before paying, which IMD already requires [R:S1]. - Make the rider's single payment a Claus escrow deal, with the coordinator as provider. The coordinator pays IMD's 0.5 IMD out of its own float, and earns it back only on acceptance. That puts the no-refund risk on the party able to judge it. - Publish a monitoring agent that disputes within the 24-hour window. ### 1h. Where the open City lets spam into swarm jobs - Wallet login is a challenge you sign with any EOA [R:S12]. - Free Jobs need no funded wallet. The board allows 1,000 memberships, 1,000 changes a day, and 200 changes per participant per day [R:S9]. - A claim locks a task for one hour [R:S9]. So a few fresh wallets can claim every new task, sit on it, and let the claims lapse. That denies service for 1 hour per claim, at no cost [I]. - Briefs (up to 4,000 characters) and outputs are public and go into agents' prompts. That is a prompt-injection channel into any seat that picks up City tasks [R:S9][I]. IMD warns the same about its own briefs [R:S1]. - Public Agent Chat is readable by anyone, and published agents can post [R:S6]. **Proposal** [P] - Swarm jobs never come from the open board. They enter only as paid IMD intake requests, or as Claus tasks that already have a funded escrow agreement. - Claims on swarm jobs require holding a seat. - Workers treat all task text as data. Both sides already instruct this [R:S9][R:S1]. --- ## 2. Five ranked changes to v0.1 **1. Make the first proof a single IMD audit of a Claus upgrade, ordered over the existing Robinhood intake, with nothing built first.** [P] - **What changes:** Claus pays 0.5 IMD per request on chain 4663 to `0xa43e…de82` for `template: "audit"` on the public repo and commit of its next token or hook upgrade [R:S1]. It publishes the report in its onchain Journal (`claus-journal/1`) [R:S6], and holds the upgrade until the report is public. - **Reason:** - It tests "IMD builds and audits, Claus operates" with zero new code and the least trust. The audit is read-only, so it "fixes nothing" [R:S1]. - It aims review at the highest-risk keys on the Claus side (F2). - It replaces v0.1's plan for Claus to coordinate upgrade review itself. That plan rests on a coordinator, seats and acceptance logic that Claus does not have (v0.1's own gap list), and on Haiku with 1,024-token outputs (F4). - **Cost:** - A few IMD (one request per upgrade, plus `job.continue` re-checks). - Claus has to wait for the report before each upgrade. - It is not independent of IMD's operator set, which is acceptable for a first proof. **2. Lock or publish the rule-changing keys before any money crosses.** [P] - **What changes:** this is a gate between phase 1 and phase 2, using the lists in §1f. - Claus: a multisig, a timelock of at least 72 hours, an onchain mint cap, and one consistent fee state. - IMD: the Intake owner and admin powers, and how the writer, quote signer and payee keys are held. - **Reason:** a posted price in CLAUS, a "stated split between burn, seats and treasury", and a seat fee share all mean nothing while one EOA can mint, re-route fees or upgrade with an hour's unenforced notice [R:S7]. On the IMD side, one writer key decides every result and failure [R:S3]. - **Cost:** - Claus loses its ability to ship upgrades quickly. The Journal says the hooks are meant to keep evolving [R:S6]. - A multisig needs signers both teams trust. - IMD has to write documentation. **3. Base cross review and the record on operators and paid, still-live work, not seats or acceptances.** [P] - **What changes:** - Use operator de-duplication and cross-holding statements (§1a, §1b). - The other swarm's coordinator picks the reviewer (§1c), and reviews are committed and revealed (§1d). - The record counts only paid, unlinked, still-live-at-90-days work, read through a published `clientAddresses` list (§1e). - No routing priority comes from the record until it has run for one full phase. - **Reason:** every gaming path in the lens goes through counting seats or acceptances. ERC-8004 itself puts the Sybil problem on readers [R:S18]. - **Cost:** - A small piece of operator-graph and commit-reveal code in each coordinator. - Honest people who hold seats on both sides get excluded from some reviews. - The record moves more slowly. **4. Reverse who reviews what in the first coordinated job, and defer Claus seats to after a Claus coordinator exists.** [P] - **What changes:** v0.1 makes "review of every upgrade to a token, hook or vault" Claus's first coordinated job. But the Claus upgrades that most need review are the ones its own single admin controls (F2). A swarm checking its own operator's upgrades is the collusion case in its plainest form [I]. So IMD (or a third party) reviews Claus upgrades, and Claus seats start with jobs where a wrong answer costs little: hosting checks and site reviews of IMD builds. - **Reason:** - Independence. - Today's Claus sponsored inference cannot do a serious contract review, and has not been shown to serve 300 seats (F4). - **Cost:** - Claus gives up the prestige job type at the start. - Houses on the holder's own plan still need the seat software, which is v0.1 piece 1. **5. Make the rider's "one payment" an escrowed Claus deal with a stated sunk IMD fee, and make the rider record read-only.** [P] - **What changes:** - The rider funds one Claus escrow (the refund path already exists [R:S10]). - The coordinator pays IMD's non-refundable 0.5 IMD and carries that risk (§1g). - The "rider record kept by both" becomes a public list of the IMD job IDs and Claus task IDs, with the payment for each, that anyone can read again. It earns no credit or fee share until change 3 has run. - **Reason:** - The two payment models have different failure rules (F6, §1g). - A record that earns money from day one invites farming (§1e). - **Cost:** - The coordinator needs a working float. - Riders see less reward early on. - Paid deals on Claus are documented on chain 1 in the examples; whether they work on Robinhood Chain is an open question (§4). --- ## 3. What I would cut - **[P] Phase 2's "pilot of 30 seats doing upgrade review"** as written. Replace it with change 1, then a pilot of Claus seats doing lower-stakes jobs (change 4). - **[P] Piece 6 (rider layer) and the swarm native company, until phase 4 or later.** They need every other piece to be trustworthy, and they add the most to the attack surface (§1e, §1g). - **[P] "Earns priority, credit and a fee share" for riders and seats in phases 1–3.** Keep the record; drop what it pays out (F9, §1e). - **[P] The two-way bridge in phase 1.** Start one way: Claus orders IMD work through the existing intake. The reverse direction, an IMD workflow ending with a paid Claus x402 call, has no published Claus x402 hosting or review service to call yet. Its x402 offers are project services, and its project inference is closed [R:S6][R:S15]. That makes it new code on both sides. - **[P] A "stated split between burn, seats and treasury" in the posted price**, until F1 is resolved and change 2 is done. - **[P] The decision on a second seat class in phase 4.** Nothing in phases 1–3 produces evidence for it. ## 4. What I would keep exactly as it is - **No shared token: each swarm is paid in its own.** This keeps either token's admin risk from spreading into the other swarm's payment rail [I]. - **Phase 0: research and a yes or no.** - **The basic division of labor: IMD builds (contracts, sites, audits, launches); Claus houses and operates.** It matches what each actually ships today [R:S1][R:S16]. - **The swarm card's six questions.** I keep them word for word, and add one required field beside them in the revision: rule-change keys and notice. - **"Work above a set value gets its second review from the other swarm."** This is the right rule, once change 3 defines "other" by operator rather than by swarm membership. - **"Still live later" as a part of the record.** It is the hardest signal to farm. ## 5. Open questions only the founders can answer 1. **Claus:** Will the Robinhood owner move to a multisig with an onchain timelock and a mint cap, and when? Which fee split is current, 1% burn and 1% treasury [R:S7], or all to liquidity [R:S6]? 2. **Claus:** Is CLAUS on Ethereum still upgradeable? [R:S6] and [R:S10] disagree. 3. **IMD:** Who owns the Intake, if anyone, and what can that owner change? How are the writer, quote signer and payee keys held: hardware, multisig, or a hot server key? 4. **IMD:** Will IMD write reputation for non-IMD agents, or accept a WorkRegistry receipt about Claus work? Or will cross-swarm records live only on the reader's side? 5. **Claus:** Is the CLAUS Jobs escrow deployed on Robinhood Chain (4663), or only on Ethereum? The docs' example uses `chainId: 1` [R:S9]. 6. **Claus:** Will house sponsorship ($5 a day, Haiku 5.5) continue once houses run on the holder's own plan, or end? 7. **Both:** Will each side accept the other's coordinator picking its reviewers for cross review? 8. **Both:** What dollar value of a job triggers cross review, and who carries the no-refund IMD fee on a cross-swarm order? 9. **Proposer:** Where is the Matt Huang post (F10)? --- ## 6. Revised design, one page (v0.2, my proposal [P]) **Goal.** Claus and IMD hire each other through what already exists. A third swarm can join by publishing a card. Trust is earned through paid, checkable work before any shared reputation pays anything out. **Principles** - Each swarm is paid in its own token. - No rule-changing key without published custody and notice. - Count operators, not seats. - Count paid, still-live work, not acceptances. **Swarm card: seven fields** The six original questions: 1. Who are your seats? 2. What do you sell? 3. What does it cost? 4. How do I pay? 5. What do I get back? 6. Who checked it? Plus one new field: (7) **which keys can change the rules, who holds them, and with how much onchain notice.** **Phases** - **0. Research, F1–F12 fixed, yes or no.** - **1. One-way proof, no new code.** Claus orders IMD audits of its next upgrades via the Robinhood intake. Each upgrade waits for its report, and the report goes into the Claus Journal. IMD publishes two or three builds as static sites in the City using Claus's existing publisher [R:S16]. - **Gate.** Claus has a multisig, a timelock of at least 72 hours, a mint cap, and one fee state. IMD has published the Intake owner and powers and its key custody. Both cards are published. - **2. Claus as a seat swarm, lower stakes.** A few dozen houses run seat software on their holders' own plans and do hosting and site review for IMD builds. Cross-holding statements are published. The other side's coordinator picks reviewers, and reviews are committed and revealed. - **3. Cross review above a set value**, run the other way where Claus has capacity. The record goes live, read-only: paid, unlinked, still live at 90 days, filtered by a published list of trusted requesters. - **4. The rider gets one escrowed payment** (a Claus deal with a coordinator float covering IMD's sunk fee). The record starts earning priority and a fee share only after a full phase of clean data. Then the first swarm native company, and a third swarm. **Out until proven:** a second seat class, priority from reputation, and a burn/seat/treasury split built into prices. **Failure rules** - Never move value on an IMD failure callback [R:S3]. - Riders' money sits in immutable escrow with a deadline refund [R:S10]. - A watcher disputes within 24 hours. - Swarm jobs never come from the free board. --- ## Sources (all read 2026-11-12, public GET only) - **S1** IMD skill: https://imd.fun/skill.md - **S2** IMD docs (full text): https://imd.fun/docs/llms.txt (sections "What a seat earns", "Reviews", "Contracts", "Wallets"; also https://imd.fun/docs) - **S3** IMD pay onchain: https://imd.fun/skill/onchain.md - **S4** IMD worker / who gets work: https://imd.fun/skill/worker.md - **S5** IMD live capabilities: https://api.imd.fun/requests/capabilities - **S6** Claus llms.txt: https://claus.si/llms.txt - **S7** Claus migration reference (contentUpdatedAt 2026-11-06): https://claus.si/migration-reference.json - **S8** Claus homepage: https://claus.si - **S9** Claus Jobs guide: https://claus.si/docs/jobs - **S10** Claus paid deals / escrow: https://claus.si/docs/paid-deals - **S11** Claus work in the network: https://claus.si/docs/agent-economy - **S12** Claus NFT inference and agent accounts: https://claus.si/docs/nft-inference - **S13** Claus live NFT capabilities: https://claus.si/__network/v1/nft/capabilities - **S14** Claus NFT collection config: https://claus.si/nft-config.json - **S15** Claus inference status: https://claus.si/docs/inference - **S16** Claus agent start guide: https://claus.si/build.txt - **S17** Claus project JSON: https://claus.si/about.json - **S18** ERC-8004 Trustless Agents (DRAFT): https://eips.ethereum.org/EIPS/eip-8004 - **S19** x402 README (schemes `exact`, `upto`, `batch-settlement`; facilitator choice): https://github.com/coinbase/x402 - **S20** IMD worker on Docker (linked from S4, not opened): https://github.com/isfinne/imd-worker-docker - **Searched, not found:** Matt Huang "swarm riders" post (web search, 2026-11-12). **Limits** - Every source is a page the project itself publishes. Neither team's claims were checked against chain state, so the contract and owner facts are as the projects state them, on the date given. - The other four views were not read.