# View 2: Claus as a swarm next to IMD. How it breaks, and what to cut Researcher 2 of 5, working alone. Lens: adversaries and failure. Date of reading: 2026-11-12. **Labels.** **[R]** = read from a cited public source. **[C]** = read directly from the chain through public read-only calls (`eth_call`, `eth_getCode`, `eth_getStorageAt`); the public RPC for Robinhood Chain is `https://rpc.mainnet.chain.robinhood.com` and for Ethereum `https://eth.drpc.org`, as given in [S2]. **[I]** = my inference. **[P]** = my proposal. **[?]** = I could not verify it. **Scope and limits.** I read public pages and made public read-only chain calls. I did not log in, post, sign or call any write or private endpoint. I did not decode every admin function of the contracts. I did not test whether the same people hold seats in both swarms (see Open questions, Q9). --- ## 0. Facts in v0.1 that are wrong or stale, corrected first | # | v0.1 says | What the sources show | Label | |---|---|---|---| | F1 | Claus has "a 2% pool fee, half to buy and burn CLAUS, half to compute and hosting" | On Robinhood the 2% project fee goes **entirely to liquidity**. There is no treasury share and the keeper adds it in batches. On Ethereum, a 2.3% fee accrues as ETH to one fixed recipient, `0xBb14…F1C5` ("CLAUS DEV"). New Ethereum NFT, LP-reward and buyback allocations are paused [S1]. The "burn plus compute" split is history, not current fact. | R | | F2 | "300 NFT houses" can become seats "under ERC-8004" | The houses are an ERC-721 collection **on Ethereum** (`0x8039…95d9`, chainId 1) [S6]. Each one holds 50,000 deposited CLAUS, and its owner can burn it to redeem them [S1]. "NFTs, backing and stakes remain on Ethereum" [S1]. All 300 are minted. 185 are staked in a contract (`0x4e59…CAD4`, which has code) and record a beneficiary. Counting beneficiaries, there are **147 distinct holders**. The largest holds 22, and the next ones hold 16, 16, 15 and 11 [S7, my count at Ethereum block 26397881]. | R/C/I | | F3 | Houses have "sponsored inference up to 5 dollars a day" (implied: enough to be a seat) | Correct that the cap is $5 per NFT per day [S8]. But the live profile is **`claude-haiku-5.5` only, at most 1,024 output tokens, 6 requests a minute and 1 at a time**, and it reports `supports300Independent: false` [S9]. IMD contract work requires `claude-fable-5-1`/`gpt-6-astra` at high effort or `claude-opus-5-5`/`gpt-6.1-sol` at medium effort, plus Foundry [S2]. A sponsored house therefore cannot do IMD-grade contract or review work. | R/I | | F4 | Claus has "no independent acceptance or onchain record" | This overstates the gap. Claus Jobs already has: a reviewer committed in `reviewPolicy` before a claim; frozen input and output hashes; one immutable report per reviewer, criterion by criterion; a rule that the reviewer cannot be the requester, the worker or their known controller; reviews paid through a separate commission; an onchain delivery commitment for paid tasks; a 24-hour review window; a 3-day dispute window; and fallback settlement [S4]. What is missing is binding acceptance (review is advisory, and only the requester accepts) and a reputation that can be read across swarms. | R | | F5 | The jobs board has "optional ETH or CLAUS escrow" (implied: on Robinhood Chain) | Every escrow example uses `chainId: 1` (Ethereum) [S4]. I found no Robinhood Chain escrow in the docs I read. A Claus-to-IMD bridge "on Robinhood Chain" would therefore cross chains on the Claus side. | R/I | | F6 | Claus has "395 API operations" | The release page (26 Oct) says **351** documented operations [S10]. The live OpenAPI, version 0.41.0, lists **426** path-and-method operations by my count [S11]. | R/I | | F7 | IMD charges "flat 0.5 IMD a request" (implied: it burns or goes nowhere) | 0.5 IMD per request is correct. The Intake "forward[s it] whole to the treasury in the same call", to a payee wallet `0x4e0f…dbc` [S2]. The same docs also say "Nobody takes a fee. There is no treasury" [S2]. The docs contradict themselves, and the founders should fix that. | R | | F8 | ERC-8004 "seats" with an onchain record on both chains | IMD's ERC-8004 adapter, WorkRegistry and ReputationRegistry are listed **only on Ethereum and Sepolia**, not on Robinhood [S2]. Every IMD review is written by one wallet, the WorkRegistry writer `0x5da3…6d44` [S2]. | R | | F9 | "Matt Huang's post of Nov 12 2026 on swarm riders" | I could not find the post. A web search turned up nothing on "swarm riders" for him [S14]. The thesis is treated here as the proposer's framing and is unverified. | ? | | F10 | "New project inference submissions switched off on Oct 21" | The docs confirm legacy project inference is disabled and separate from NFT sponsorship [S1]. I did not confirm the exact date. | R/? | --- ## 1. Answers to the lens questions ### 1a. Who can change the rules (keys) **IMD Intake v2** (`0xa43e…de82`, the same address on Ethereum, Robinhood and Sepolia) [S2]: - `owner()` returns `0x047f606fd5b2baa5f5c6c4ab8958e45cb6b054b7` on both Ethereum and Robinhood. That address has **no code**, so it is a plain key (EOA), not a multisig [C]. - `writer()` returns `0x3d3c…25cb`, which matches the docs and is also an EOA [C]. The docs say the writer key "cannot forge an answer but can withhold one, pair one with the wrong request … or send a false failure" [S2]. - The EIP-1967 implementation and admin slots are empty. The contract also exposes `transferOwnership` (selector `f2fde38b` appears in its bytecode) [C]. I infer it is not an upgradeable proxy; the docs say new versions land at new addresses [S2]. **[I]** What the owner can change (price, payee, writer, sold actions) is not documented in what I read. I did not decode it. - Other single keys named in the docs: the quote signer `0x169b…279f`, the launch deployer `0xcecc…a551`, and the WorkRegistry writer `0x5da3…6d44` [S2]. Schedule owners "have no controls: pausing and cancelling stay with the IMD team" [S2]. **Claus Robinhood token** (`0x1630…EfFAC`) [S1]: - It is an EIP-1967 proxy. Implementation `0xd687…4bff`, ProxyAdmin `0xcbc4…22a8` [C]. - The token's `owner()` and the ProxyAdmin's `owner()` are both the contract `0x979d…dc46`. That contract's `owner()` is `0xbb14…f1c5`, which has **no code on Robinhood** [C]. On Ethereum the same address carries an EIP-7702 delegation designator (`0xef0100…`) [C]. It is the same address as the fixed 2.3% fee recipient [S1]. **[I]** One key controls upgrades of the token, hook and liquidity management, and the mint. - The docs say so too: "owner-upgradeable … a single admin wallet with extra mint authority and no onchain mint cap or timelock … at least 1 hour of notice … not enforced by the contracts" [S1]. - The Ethereum CLAUS token is different: its ProxyAdmin was renounced and only metadata control remains [S1]. **What each side should publish or lock before money crosses [P]:** 1. **Claus:** move the token, hook and ProxyAdmin ownership behind a timelock of at least 48 hours, owned by a multisig with named signers. Put a hard onchain mint cap. Otherwise CLAUS should not be the unit for posted prices or seat payments: one key can mint or change fees under any price the swarm card states. Cost: one deployment plus an ownership transfer, and each later upgrade takes 48 hours. 2. **IMD:** publish what the Intake owner can call. Move the owner from an EOA to a multisig. Publish a rotation and compromise plan for the writer and quote-signer keys. Cost: low, mostly docs and one `transferOwnership`. 3. **Both:** publish one signed "keys file" per swarm, as a section of the swarm card, listing every privileged address, what it can do, and its timelock. A consumer can then check it against the chain, as I did above. ### 1b. Fake or duplicate seats, and one owner in both swarms - **IMD:** 2,000 NFTs, minted "free, one per wallet" [S2]. They are transferable, so the one-per-wallet rule at mint says nothing about who holds them today. Pairing refuses a device that already serves another seat (`device_serves_another_seat`) [S2]. Nothing stops one person running many devices. "Nobody reviews their own work" is enforced per seat [S2]. **[I]** It does not stop a holder of seats A and B from building on A and reviewing on B. - **Claus:** the docs state the limit outright: "Separate identities do not prove separate real-world operators" [S4]. Holdings are already concentrated, with the top five beneficiaries holding 80 of 300 (F2). **[I]** In a 30-seat pilot drawn from these holders, a few wallets could fill most of the review panel. - **Cross-swarm:** v0.1 says work above a set value gets its second review "from the other swarm". **[I]** That defeats itself if the same person holds the IMD builder seat and the Claus reviewer house. Nothing in either system links the two. - **Cheapest defence [P]:** at assignment time, exclude any reviewer whose current holder wallet (for Claus, the staking beneficiary) holds or benefits from any seat that touched the job in either swarm. Both holder lists are public and cheap to read (S7; `ownerOf` on `0x0000ec93…ec1d`). This is a read and a set intersection. Leave funding-graph checks for later. ### 1c. Reviewer collusion and copied reviews - **IMD's gate is mechanical.** The verifier replays the signed bundle and runs `forge build`/`forge test`. Adversarial review blocks only on "a high or critical finding with a reproduction" [S2]. **[I]** Collusion can hide findings by staying silent, but it cannot pass failing tests. - **Claus review is advisory prose against criteria**, bound to output hashes [S4]. **[I]** A colluding reviewer can return "pass" on every criterion at no cost. A copier can paste another reviewer's rationale and only needs the hashes. - **Cheapest defences [P]:** - The coordinator assigns reviewers at random from the eligible set. The requester does not pick them; today Claus lets the requester choose [S4]. - Count a review as an upgrade review only if it carries a runnable artifact: a failing test, a fork-replay script, or a diff reproduction. - Seed known-bad canary upgrades, about 1 in 10. A reviewer who passes a canary loses record weight. - Commit and reveal for multi-reviewer panels. Each reviewer submits a hash of the report before any report is visible, which kills copying. Claus already stores report hashes [S4], so this is a small change. ### 1d. Reputation farming against the rider and seat record - **IMD's score.** Under `acceptance-v2`, each accepted node scores 1 and each rejected node 0. Past wins do not affect dispatch [S2]. The 2% work share goes to "wallets with accepted work on that launch", and a launch costs 0.5 IMD [S2]. **[I]** A seat holder can pay 0.5 IMD for trivial jobs they hope to win. Dispatch is ranked by tools, load, failures and turn, not by the holder [S2], so winning is not guaranteed. Cheap launches would still let a group farm both accepted reviews and the 2% share. - **Claus.** The free board allows 200 changes per participant per day and needs no funded wallet [S3]. **[I]** Self-posted and self-accepted tasks between two wallets one person controls produce an "accepted" portfolio at zero cost. `public_agent_portfolio` lists accepted tasks [S12], and the docs themselves warn that "application acceptance, independent counterparties and payment are separate facts" [S12]. - **ERC-8004 itself** says reads without a `clientAddress` filter "are subject to Sybil/spam attacks" [S5]. - **Cheapest defences [P]:** - The rider and seat record counts only **paid** jobs whose payer has no seat link to the worker. - Weight each job by value paid and by distinct payers. - The "still live later" part of the record is read from chain at day 30 and day 90, not reported. - Consumers read ERC-8004 summaries filtered to an allowlist of client addresses, meaning each swarm's published WorkRegistry writer. ### 1e. A paid job that returns nothing - **IMD:** no refunds. A refused body is a spent payment. Status 2 (`failed`, `disagreed`, `blocked`, `cancelled`…) comes with no result, and failures are **unsigned** ("anyone can trigger one by asking with a bad body"). Results arrive "minutes to days" later "and can be late or missing" [S2]. `/requests/check` and `/requests/quote` are free and validate the body before payment [S3-skill]. - **Claus:** an escrow paid task has `refund`, `settle`, a 24-hour delivery review window, a 3-day dispute window and fallback settlement [S4]. Free tasks give nothing back. - **Failure mode [I].** In the bridge, Claus pays 0.5 IMD into a black box with no recourse. In the other direction, IMD pays a Claus x402 service where settlement is immediate. x402 is "trust minimizing" about the facilitator moving funds, but the core flow has no refund step [S13]. Neither direction has the same guarantees as Claus's own escrow. - **Cheapest defences [P]:** 1. Every bridge order runs `check` and `quote` first. 2. Claus orders through a contract based on the documented `IntakeClient`, with a hard per-day spend cap. On failure it clears the pending request and never moves value [S2]. 3. For IMD buying from Claus, use the existing **Claus Jobs escrow** with `reviewPolicy`, not a bare x402 call, so payment releases only on acceptance or after the deadline. 4. Ask IMD for one policy change: a job that ends `blocked` with no deliverable gets a free `job.continue` credit. This is a founder decision (Q4). ### 1f. Where the open City lets spam into swarm jobs - **What the board allows [R].** Wallet login, free tasks with no funded wallet, briefs up to 4,000 characters, 5,000 stored records, 1,000 changes a day board-wide and 200 per participant [S3]. - **Board exhaustion [I].** About five fresh wallets can use up the board's daily change budget. 5,000 records can be filled, which is a denial of service against real work. - **Prompt injection [I].** IMD bodies "go into the builders' prompts" and are public forever [S3-skill]. Any route that turns a free Claus task into an IMD job lets anyone inject prompts into IMD seats. - **Cheapest defence [P].** Only paid or escrowed tasks may become swarm jobs. The free board never feeds the coordinator. Each cross-swarm brief passes IMD's free `requests/check` before payment. ### 1g. What to cut or reorder for a first proof with the least trust and code See sections 3 and 6. In short: run **Phase 1 only**, use **existing rails only** (IMD Intake plus the `audit` template; Claus Jobs escrow plus `reviewPolicy`), and make the first job **IMD auditing a Claus hook upgrade**. Defer houses-as-seats, the coordinator, the CLAUS price and the rider layer. --- ## 2. The five changes I would make to v0.1, ranked **1. Make a key lock-down a precondition, before any money crosses. [P]** - *Reason:* the Claus Robinhood token, hook and liquidity are upgradeable and mintable by one EOA with no timelock [S1][C]. IMD's Intake owner and writer are EOAs [C]. A posted price "in CLAUS" or a seat payment in CLAUS is only as good as that key. Making "upgrade review" the first job while one key can skip review defeats its own purpose. - *Cost:* a timelock deployment and a multisig on the Claus side; slower upgrades, at least 48 hours; one ownership transfer and documentation on the IMD side. No new product code. **2. Do not turn the 300 houses into seats yet. Recruit Claus reviewers from people who run their own frontier plan. [P]** - *Reason:* the houses are Ethereum NFTs whose sponsored inference is Haiku-only, 1,024 tokens and 6 requests a minute, and cannot serve 300 at once (F2, F3). The holdings are concentrated (F2). Making them seats either buys low-grade reviews or leaves the treasury paying for them, which is the "inference paid by the treasury" gap v0.1 itself names. A Claus reviewer should be any wallet with a Claus identity that runs its own model, like IMD's worker rule [S2]. - *Cost:* the house holders lose an expected perk, which is politically hard. No code. **3. Make cross review mechanical and conflict-checked, not prose. [P]** - *Reason:* sections 1b and 1c. Per-seat self-review checks do not cover one person in two swarms. Claus reviews can be copied and colluded on. - *What to require:* random assignment; exclusion by holder or beneficiary across both swarms' public holder lists; a runnable artifact; commit-and-reveal; canary jobs. - *Cost:* a small coordinator script (read two holder lists, draw at random) and a canary budget, about 10% more jobs. **4. Count only paid, unlinked and later-verified work in any record. [P]** - *Reason:* section 1d. Free, self-dealt or cheap jobs farm acceptance, and ERC-8004 is Sybil-prone when reads are unfiltered [S5]. - *What to do:* each swarm publishes its WorkRegistry writer as the only trusted client address, and readers filter by it. The "still live" checks are chain reads at 30 and 90 days. - *Cost:* the record grows slowly, and early riders have thin records. **5. Use what exists for refunds and payment direction; add no new payment primitive. [P]** - *Reason:* section 1e. IMD has no refunds and unsigned failures. Claus Jobs already has escrow, a review window, disputes and fallback. - *What to do:* IMD buys from Claus through Claus Jobs escrow. Claus buys from IMD through a capped `IntakeClient` contract after `check` and `quote`. Drop "one posted price per request in CLAUS over x402" until change 1 is done. - *Cost:* the CLAUS-side burn/seat/treasury split waits. The Claus escrow is on Ethereum (F5), so Claus pays Ethereum gas. ## 3. What I would cut - **Piece (1), houses become seats**: cut for now (change 2). - **Piece (4), one posted CLAUS price with a burn/seat/treasury split**: cut until change 1. The fee facts it assumes are already wrong (F1). - **Piece (6), the rider layer** (one brief, one payment, a shared rider record): cut from v0.1. It needs a joint coordinator, a refund policy and a farm-resistant record, and none of these exist yet. It is the most new code and the most trust. - **Phase 4's "first swarm native company"**: cut. A company with "no staff" whose contracts sit behind single EOA keys is the worst case for this lens. - **Phase 2 at 30 seats**: shrink to 5 to 10 named reviewers on their own plans, across at least 5 distinct beneficiary wallets. - **"Work above a set value gets its second review from the other swarm"**: hold it until the cross-swarm conflict check (change 3) exists. Without the check it is worse than a single-swarm review. ## 4. What I would keep exactly as is - **No shared token; each swarm is paid in its own** [P keep]. It limits contagion when one side's keys fail. - **The swarm card's six questions** [P keep], with one addition I treat as part of "who checked it": a keys section (1a). - **Upgrade review as the first coordinated job type** [P keep]. It is the job where an independent second swarm adds the most, because IMD already sells `audit` with `repoUrl`/`baseCommit` [S3-skill]. - **Phase 0, a yes or no, and Phase 1, hire each other with what exists** [P keep]. This is the cheapest proof. - **The split of roles: IMD builds and audits, Claus hosts and reviews** [P keep]. ## 5. Open questions only the founders can answer 1. **Claus:** will you move token, hook and ProxyAdmin ownership (`0x979d…`, owned by `0xbb14…`) to a timelocked multisig and cap the mint before any CLAUS-priced swarm work? Who would the signers be? 2. **IMD:** what can the Intake owner `0x047f…54b7` change (price, payee, writer, sold actions)? Will it move to a multisig? 3. **IMD:** which is true, "forwarded whole to the treasury" or "there is no treasury" [S2]? Where does the 0.5 IMD go? 4. **IMD:** will a job that ends `blocked` or `failed` with no deliverable earn a free continuation or credit for a cross-swarm buyer? 5. **Claus:** will NFT sponsorship stay Haiku at $5 a day, or would you fund a reviewer tier on a frontier model? If neither, do you accept that reviewers run on their own plans? 6. **Claus:** is there, or will there be, a Robinhood Chain escrow for Jobs, or is Ethereum the only settlement chain? 7. **Both:** who runs the cross-swarm reviewer assignment, and who can override it? 8. **Both:** who sets the "set value" above which cross review is required, and in which unit? 9. **Both:** may we publish, once, the overlap between IMD seat holders (`0x0000ec93…`) and Claus house beneficiaries, so the conflict check starts from a known baseline? I did not finish this check. 10. **Proposer:** can you share the Matt Huang post? I could not find it (F9). --- ## 6. Revised one-page design (v0.2, my proposal) [P] **Goal.** Prove that two independent swarms can buy and check each other's work with the least trust and no new contracts. Then publish a swarm card any third swarm can copy. **Preconditions, published before money crosses:** - (a) Each swarm publishes a keys file listing every privileged address, its powers and its timelock. - (b) Claus moves token, hook and ProxyAdmin ownership to a multisig behind a timelock of at least 48 hours and caps the mint, or else nothing is priced in CLAUS. - (c) IMD documents the Intake owner's powers and moves the owner to a multisig. **Phase 0: research and a yes or no.** Correct the facts (section 0). Agree the conflict rule and the canary rate. **Phase 1: hire each other with what exists. No new code.** - **Claus to IMD:** Claus orders 3 IMD `audit` jobs on its own pending hook or token upgrades through Intake v2 on Robinhood, from a capped `IntakeClient` contract, after the free `check` and `quote`. - **IMD to Claus:** IMD posts 3 review or hosting tasks on Claus Jobs with escrow and a `reviewPolicy`, and publishes 3 builds in the City. - **Success:** 6 jobs delivered, each with public hashes; any finding reproduced; zero disputes unresolved after the 3-day window. **Phase 2: a pilot of 5 to 10 reviewers.** Claus reviewers run on their own frontier plans, with at least 5 distinct beneficiary wallets. Assignment is random, with holder exclusion across both swarms. Each review needs a runnable artifact; reports use commit and reveal; 10% of jobs are canaries. **Phase 3: the swarm card and the record.** Each swarm answers six questions: seats, what it sells, price, how to pay, what you get back, who checked it, plus its keys. The record counts only paid, unlinked jobs. It is read from each swarm's own WorkRegistry writer through ERC-8004 summaries filtered by client address, with chain checks of "still live" at 30 and 90 days. **Phase 4: decide on the deferred pieces.** Houses as a seat class; a CLAUS price with its split; a rider layer; a swarm native company. Each is decided only if Phases 1 to 3 met their numbers and the preconditions still hold. Open the card to a third swarm. **Out of scope for v0.2:** a shared token, a shared coordinator, refunds beyond what each rail already has, and free tasks feeding swarm jobs. --- ## Sources - [S1] Claus, llms.txt (migration, fees, upgrade authority, NFTs): https://claus.si/llms.txt - [S2] IMD docs, full: https://imd.fun/docs/llms.txt (seat earnings, Who gets work, Reviews, Intake, contracts and wallets tables) - [S3] Claus Jobs: https://claus.si/docs/jobs. [S3-skill] IMD skill: https://imd.fun/skill.md - [S4] Claus Jobs, "Independent task review" and payments sections: https://claus.si/docs/jobs - [S5] ERC-8004 Trustless Agents (draft), reputation and security considerations: https://eips.ethereum.org/EIPS/eip-8004 - [S6] Claus NFT config: https://claus.si/nft-config.json - [S7] Claus live collection: https://claus.si/nfts.json (read 2026-11-12T09:12Z, Ethereum block 26397881) - [S8] Claus NFT inference: https://claus.si/docs/nft-inference and conditions https://claus.si/docs/nft-inference-terms - [S9] Claus NFT capabilities: https://claus.si/__network/v1/nft/capabilities - [S10] Claus release status (API 0.40.0, 26 Oct 2026): https://claus.si/docs/release-status - [S11] Claus OpenAPI 0.41.0: https://claus.si/__network/openapi.json - [S12] Claus "Work in Claus Network": https://claus.si/docs/agent-economy - [S13] x402 README: https://github.com/coinbase/x402 (raw: https://raw.githubusercontent.com/coinbase/x402/main/README.md) - [S14] Web search "Matt Huang swarm riders swarm native companies" (2026-11-12): no matching post. Example results: https://www.paradigm.xyz/team/matthuang, https://www.vcsheet.com/who/matt-huang - [C] Chain reads through the public RPCs named in [S2] (Robinhood `https://rpc.mainnet.chain.robinhood.com`, Ethereum `https://eth.drpc.org`): - IMD Intake `0xa43e6f75ee006411f79ac1c84120606c2330de82`: `owner()` returns `0x047f606fd5b2baa5f5c6c4ab8958e45cb6b054b7`; `writer()` returns `0x3d3c0df7c1a17d384545e9d7c8228897bd9625cb`; both have empty code; the EIP-1967 slots are empty. - CLAUS Robinhood token `0x163015276f3244bC2CcB78b4D69cc8A4069EfFAC`: EIP-1967 implementation `0xd68727ddd50e8ba07e4be20d37a5ecf3dd814bff`; admin `0xcbc4bbd337e62c926c995e66a231e5e2b0a722a8`; token `owner()` and admin `owner()` are both `0x979d5810d32292358639f25294472f01c95bdc46` (a contract); its `owner()` is `0xbb145ca83272c3806d4ddc75ca1d5514789cf1c5`, with empty code on Robinhood and an EIP-7702 designator on Ethereum. - Robinhood block about 0x6b34804; Ethereum block about 0x192ccb7 at the time of reading.