# View 4 — Claus as a swarm adjacent to IMD: how it breaks, what to cut Lens: adversary and failure. Date of reading: 2026-11-12. Desk research only; no login, no probe, no request to any write endpoint (only plain page GETs). **Tags used:** **[R]** read from a source (cited). **[I]** my inference. **[P]** my proposal. **[U]** unknown or unverified. ## 0. Sources and their limits | # | Source | What it gave | |---|---|---| | S1 | https://claus.si (home page) | CLAUS token on Robinhood Chain (4663) at `0x163015276f3244bC2CcB78b4D69cc8A4069EfFAC`. Fee on each pool trade, part buys and burns, rest funds compute, hosting, work. The page says the token, hook and vault "can be upgraded around the same address" and does not say who controls upgrades. Old Ethereum token retired, snapshot block 26357823. | | S2 | https://claus.si/Jobs and https://claus.si/Hooks | Rendered only navigation text. **Nothing on job rules, fee percentages, admin keys, timelock or audits could be read.** | | S3 | https://imd.fun (home) | Nav only; points agents to `imd.fun/skill.md`. | | S4 | https://imd.fun/skill.md | 0.5 IMD per request; "There are no refunds: a refused body is a spent payment"; "Payment buys the work being admitted, not a guaranteed result"; a disagreeing oracle panel ends with no answer and the payment is still spent. `POST /requests/check` and `/requests/quote` are free; a quote expires after 10 minutes. Worker needs an IMD NFT, a little ETH, and a paid Claude or Codex plan. Intake contract `0xa43e6f75ee006411f79ac1c84120606c2330de82`. **Owner and writer not stated.** Callback is a request argument (example uses zero address and zero selector). Launch fee split not given. | | S5 | https://imd.fun/docs/ | 2,000 seat NFTs; seat contract `0x0000ec93127baa929e58e97dd0095a2bfb38ec1d` on Ethereum. History says NFTs were **minted free, one per wallet, in May 2026**, while setup says "Buy one here". "The verifier checks the result before it goes onchain under your agent", criteria not described. Every accepted job adds a review under the agent. 8% of each launched token to seats connected at admission, 2% to wallets with accepted work. "Nobody takes a fee. There is no treasury." Nothing on proxies, upgrade authority, intake owner, refunds. Verifier runs Foundry v1.8.3. | | S6 | https://eips.ethereum.org/EIPS/eip-8004 | Status **Draft**. Identity registry is ERC-721. Reputation registry: any address may `giveFeedback()` except the agent's owner or approved operator. Spec states Sybil attacks are possible; `getSummary()` needs an explicit client-address list because unfiltered results are Sybil-prone. Validation registry records 0–100 responses from validators; incentives and slashing are left to the validation protocol. It "cannot cryptographically guarantee" capabilities are functional. | | S7 | https://docs.x402.org/core-concepts/http-402 | 402 challenge, signed retry, `PAYMENT-RESPONSE` reports settlement. `exact` "typically settles immediately". **Refund or escrow, and what happens when the resource returns nothing after paying, are not covered.** | Not read and **not verifiable by me [U]**: Matt Huang's post of 12 Nov 2026 (the thesis in v0.1), the Claus jobs-board rules, mandate terms, house/sponsored-inference terms, the 2% pool-fee split, the Oct 21 inference switch-off, the "395 API operations", IMD's x402 reference and callback reference, and any contract source. Every v0.1 statement about those is therefore taken as the proposer's claim, not confirmed. Several conclusions below are conditional on them. ## 1. Fact corrections to v0.1 (first) 1. **"2,000 seats … each on the holder's own machine."** [R S5] seats are NFTs that were minted free, one per wallet. One-per-wallet is not a scarcity or identity control. [I] Seat count is a count of wallets, not people or machines. Anything that treats a seat as a distinct reviewer is resting on this. 2. **"Seats under ERC-8004."** [R S6] ERC-8004 is a Draft, and its reputation registry explicitly does not solve Sybil. [R S5] says accepted work is recorded onchain "under your agent" but the docs I read do not say which registry, or whether it is the draft ERC-8004 contract. [U] 3. **"A stated split between burn, seats and treasury."** [R S5] IMD says "no treasury" and job payments are not split (removed from supply). Claus has a treasury (inference paid from it, per v0.1). So the split is asymmetric by design; the card must say so, not hide it. [I] 4. **"Claus: acceptance only by requester, payment a separate step."** [U] I could not read Claus job rules. Treat as the proposer's claim. 5. **"Same intake contract on Ethereum and Robinhood Chain."** [R S4, via the page's payment commands] one address listed for both. [I] Same address on two chains does not mean same owner or same code; check each. 6. **Seats' work is "checked by other agents and tests."** [R S5] the verifier's criteria are not documented. v0.1 builds "independent second run" on top of an acceptance rule nobody outside can inspect. ## 2. The five changes, ranked ### Change 1 — Replace "independent second run + reviewers, written to the seat record" with a deterministic-first acceptance and a *named* reviewer set (cut collusion/Sybil at the root) - **Reason.** Every attack in the lens (fake seats, one owner on both sides, collusion, copied reviews, farming) lands on the same weak point: a review is cheap to produce and a seat is cheap to mint [R S5: free, one per wallet; R S6: Sybil possible, any non-owner can give feedback]. A cross-review where reviewers are "other seats" is only as strong as seat identity. [I] - **Do.** [P] For upgrade review (a token, hook or vault upgrade) make the primary check *mechanical*: reproducible build, bytecode/storage-layout diff, Foundry tests, upgrade-authority diff, all emitted as a hash-pinned report anyone can re-run. Human/agent opinion is secondary. Reviewers are drawn from an allowlist the founders publish (start: ≤10 addresses), consumers pass that list to `getSummary()` as the spec requires [R S6]. Review text is committed (hash) before any other reviewer's is revealed, to stop copying. A reviewer address may not be funded by, or share an owner with, the author or the other swarm's reviewer (declared + checkable onchain funding graph). - **Cheapest defence per attack** (all [P] unless marked): - *Fake/duplicate seats*: do not count seats; count allowlisted reviewers. Cost: a list, zero code. For payout shares, weight by held seat *and* requirement of a live worker heartbeat, not by NFT count. - *One owner in both swarms*: [I] cannot be prevented onchain (wallets are free). Defence is to make cross review *not depend on seat ownership*: the other swarm's reviewer must be on the published allowlist and must not have a funding link to the author. Also disclose, don't ban, same-owner holdings. - *Reviewer collusion*: pay a reviewer only when a later, independent signal agrees (e.g. a bug found after acceptance, or a bounty hit), and let any allowlisted party post a dispute that revokes the record. ERC-8004 lets clients revoke and anyone append a response [R S6], so a dispute needs no new contract. - *Copied reviews*: commit-reveal hash plus similarity check on the report; a review whose hash or normalised text equals an earlier one earns nothing. - *Reputation farming*: score only work above a minimum value, weight by value paid and by age ("still live later", already in v0.1), cap per-requester contribution to one rider's score, and never let a score be earned from a counterparty the rider funds. - **Cost.** Founders must curate a list and accept it is centralised at the start. Slower to decentralise. Small code (report format + hash commit). ### Change 2 — Lock and publish the control keys *before* any money crosses (gate on Phase 1) - **Reason.** [R S1] the CLAUS token, hook and vault "can be upgraded around the same address" and the home page names no controller. [R S4/S5] the owner and writer of IMD's intake contract are not stated in the docs I read. A cross-swarm bridge that pays into either is only as safe as the key that can change it. A reviewer approving an upgrade while the reviewing swarm's own upgrade key is unpublished is incoherent. [I] - **Do (each side, [P]).** Publish in the swarm card: for every contract, the admin/owner/writer address, whether it is a multisig (signers, threshold), any timelock (delay), and whether upgradeability is renounced or scheduled. Minimum bar before money crosses: upgrades behind a timelock (≥48h, my number, founders to set), and an onchain event on every change so the other swarm can pause bridging. For IMD: state who the intake *writer* is (the party who can mark/record work) and that it cannot move user funds or edit past records. For Claus: state the proxy admin for token, hook and vault, and whether the pool-fee destination can be changed. Each side's bridge contract/caller gets a per-job and per-day cap and a pause the *receiving* side holds. - **Cost.** Founders give up instant-change flexibility; a multisig/timelock set-up is real operational work (days, not weeks). If either side refuses, Phase 1 cannot honestly start. [I] ### Change 3 — Make "paid job returns nothing" a defined outcome: escrow-then-release on the Claus side, quote-then-pay with pre-check on the IMD side, and no automatic cross-swarm retries - **Reason.** [R S4] IMD has no refunds and a disagreeing panel ends with no answer but still keeps the payment. [R S7] x402 documents the payment handshake but not recourse when the resource returns nothing; `exact` settles immediately. [v0.1] says Claus payment is a separate step after acceptance. So the same brief could be unrecoverable at IMD and unpaid at Claus. [I] A bridge that chains them (Claus orders IMD jobs; IMD ends with a paid Claus call) can drain a rider in the gap: IMD takes 0.5 IMD, returns nothing, Claus's follow-up call is made anyway. - **Do.** [P] 1. The rider's single payment is held in a Claus-side escrow (the optional ETH/CLAUS escrow already exists per v0.1, [U] unverified) and only the *per-request* IMD fee is released outward, only after IMD's free `/requests/check` and `/requests/quote` pass [R S4: both free, quote valid 10 minutes]. 2. State in the card, in plain words, per stage: "refundable / not refundable / refundable only if…". IMD stage is "not refundable"; the rider sees it before paying. 3. A downstream paid call (host the build, request second review) runs only if the upstream returned an accepted artifact; no result, no onward spend. 4. Cap IMD spend per brief; the rider is told the worst-case loss up front. If a job returns nothing the rider loses at most the capped IMD fees, not the whole escrow. 5. Record "returned nothing" in the rider record as a neutral event, not as the rider's failure and not as the seat's reputation hit unless the cause is attributable. - **Cost.** The bridge cannot be fully atomic (IMD's no-refund rule is IMD's, [I] unlikely to change). The rider carries bounded loss. Needs the escrow contract on the Claus side to be trusted, which ties back to Change 2. ### Change 4 — Reorder the phases: "prove acceptance and key hygiene with no money crossing" first; cut the pilot of 30 seats until a manual pilot works - **Reason.** v0.1 Phase 1 already moves value (three IMD jobs ordered by Claus) before any key, refund or acceptance question is settled. The least-trust first proof is one that needs no new contract and no cross-swarm payment. [P] - **Do.** New order: - **P0 (research + yes/no)**: keep, add the key publication of Change 2 as a gate. - **P1 (manual, one direction, no bridge)**: *Claus reviews one real IMD-built upgrade* (or an IMD-built hook), result published as a signed report + commit-reveal reviews from the allowlist (Change 1). Payment by hand, small and capped. This is "review of every upgrade" at n=1. Use existing x402 services and existing job board. Cost: zero new code. Success test: the report can be re-run by a stranger and gives the same diff. - **P2 (the reverse direction, three jobs)**: Claus orders IMD jobs via the existing intake contract with capped value and the pre-check of Change 3. Three IMD builds hosted on Claus. - **P3 (the swarm card)**: publish the card as a file with the six answers *and* the key table of Change 2. Only then add the rider record. - **P4 and later**: seat conversion, pilot, native company: only after P1–P3 produce a record that survived a hostile retrospective (does an outsider agree the accepted jobs were good a month later). - **Cost.** Slower to a headline number of seats. Less impressive roadmap; far less to lose. ### Change 5 — Lock the open-City intake: make a job in a swarm a *staked, rate-limited, allowlist-routed* object, not an open post - **Reason.** [R S1] Claus is an open City with wallet login; v0.1 says the jobs board is post / claim for one hour / deliver / only the requester accepts. [U] I could not read the board rules. [I] With free wallets, spam enters at four points: (a) junk jobs posted to clog claims (a 1-hour claim lock is itself a denial tool: claim everything and let it expire); (b) fake deliveries to a requester who is the attacker's own second wallet to farm "accepted" records; (c) prompt-injection in job text aimed at an agent worker or reviewer (the job body is untrusted data); (d) wash-trading of paid services to inflate rider volume. - **Do.** [P] - Only jobs of the one declared type (upgrade review) are routable to swarm workers; everything else stays a normal open-board job and never touches the rider record. - A claim requires a small refundable bond that is lost on expiry without delivery (turns the 1-hour lock from free denial into a cost). Per-wallet open-claim cap. - A requester who is also the deliverer (same wallet or a funding-linked wallet) cannot accept toward reputation; "only the requester accepts" is not enough to bear weight in a public record. Acceptance for the rider record needs Change 1's independent check. - Job text is passed to agents as quoted data with fixed output schema; agents hold no write keys beyond their mandate (30-day mandates per v0.1; [U] terms unread). - Minimum job value for a record to count, which kills dust farming. - **Cost.** Small friction on honest users; a bond contract or a bond in the existing escrow; needs the board's owner to add rules ([U] whether the board allows it). ## 3. What I would cut - **Piece (1), converting all 300 houses into ERC-8004 seats on the holder's own plan (for now).** [R v0.1] Claus's gap list says there is no working seat or coordinator; [I] this is the biggest new code and the biggest Sybil surface (300 free-to-split identities, each paying for its own inference means each also *lowers* quality if holders run weak models). Cut to a handful of operator-run reviewer seats on the allowlist (Change 1). Revisit after P3. - **Piece (6), the rider layer (one brief, one payment, a record kept by both), until P3.** It is the product *thesis* of v0.1 but also the most abusable object (reputation farming) and the largest unproven integration. Keep the idea; build it last, on top of a record that has already survived. [P] - **Phase 2 pilot of 30 seats and Phase 4 "all 300 seats / second seat class / first swarm native company".** The native company has no owner-less staff by construction ("no staff", rider owns it); [I] a legal and key-control problem (who holds the upgrade key of a company nobody employs?) that cannot be answered before Change 2. Cut from the first proof, not the roadmap forever. - **"Work above a set value gets its second review from the other swarm" as an automatic rule.** Keep as a manual option until reviewer independence (Change 1) is shown to hold; otherwise it sends the highest-value jobs to the weakest-checked reviewers. [P] - **"No shared token" is fine; cut anything that tries to reconcile prices.** Do not post one CLAUS price for IMD-fulfilled work until the exchange path is defined. [I] ## 4. What I would keep exactly as is - **Division of labor:** IMD builds, Claus houses and reviews. [I] It matches what each side can actually do from the public pages (IMD: contracts, launches, verifier, paid requests [R S4/S5]; Claus: hosting, apps, services in USDG [R v0.1, U for details]). - **Upgrade review as the first coordinated job.** It is concrete, checkable mechanically, and directly relevant to the keys issue in S1. [I] - **"No shared token; each swarm is paid in its own."** Avoids a pricing and custody dependency. [I] - **Swarm card with six questions.** Keep verbatim; add only the key table (Change 2) as an appendix, not a seventh question. [P] - **Phase 0 as research and a yes/no.** Keep. - **Records that include "still live later"** (not just "accepted"): the best anti-farming idea in v0.1, because it costs the farmer time. [I] - **IMD's free `check` and `quote` before paying** [R S4]: keep as the mandatory first step of any bridge call. ## 5. Open questions only the founders can answer 1. **Claus:** who holds upgrade authority over the token, hook and vault? Multisig or single key, timelock length, any plan to renounce? (Not stated on S1.) 2. **IMD:** who are the intake contract's owner and writer on each chain; can the writer change a recorded outcome or touch funds? Is it the same owner and code on Ethereum and Robinhood Chain? 3. **IMD:** what does the verifier actually check before an accepted record goes onchain, and can an outside party re-run it? (S5 gives no criteria.) 4. **IMD:** which chain and registry holds the seat "reviews"? Is it the draft ERC-8004 registry, or a custom contract? Can Claus read and write to it, and can the record be read from Robinhood Chain? 5. **IMD:** would you accept any refund-like path for a bridge-originated request (for instance a refused body at the pre-check costing nothing), or is "no refunds" absolute? 6. **Claus:** what exactly can a job escrow do today — does acceptance release funds automatically or is it manual? Who arbitrates a dispute? What stops claim-and-sit? 7. **Claus:** are the 300 houses transferable, can the same holder also hold IMD seats, and would you accept a published list of reviewer addresses as the first reviewer set? 8. **Both:** who is the first allowlisted reviewer set and who may change it? 9. **Both:** is there a shared willingness to publish the key table and a maximum cross-swarm value per job/day before Phase 1? 10. **Both:** what is the legal owner of a "swarm native company" with no staff, and who holds its upgrade key? 11. **Verification:** I could not read Matt Huang's post (the premise) nor the Claus jobs, hook and fee pages; please supply them, because several points above are conditional on them. ## 6. Revised one-page design (v0.2, my proposal [P]) **Goal.** Claus joins IMD as a second swarm, with a first proof that needs the least trust and the least new code. A third swarm can join the same way. **Step 0 — publish before paying.** Each swarm publishes a card: the six answers *plus* a key table (every contract, who can change it, multisig or not, timelock, pause, per-job and per-day caps). The bridge does not open until both are published and each side can pause the other's intake. **Step 1 — one real review, by hand, no bridge.** Claus reviews one real IMD-built upgrade. The primary check is mechanical and reproducible (build, test, bytecode and storage diff, upgrade-authority diff) and emitted as a hash-pinned report. Reviews come from a short published allowlist, in commit-reveal form. Pay a small capped amount over existing x402/job-board rails. **Step 2 — reverse direction, three jobs.** Claus orders three IMD jobs through IMD's existing intake contract with a hard spend cap. Each is pre-checked with IMD's free check and quote. The rider's money sits in Claus-side escrow and only the capped per-request fee leaves. A downstream paid call runs only on an accepted result. Three IMD builds are hosted in the City. **Step 3 — the record.** Acceptance written to the record requires the independent check, not just requester acceptance. Scoring weights value, age ("still live later") and a per-counterparty cap, counts only allowlisted reviewers (as ERC-8004's `getSummary` requires), and allows disputes that revoke. A returned-nothing event is recorded neutrally. **Step 4 — then, and only if Steps 1–3 survive a month of outside scrutiny:** open reviewer seats beyond the allowlist, convert houses to seats, add the rider layer, and consider a swarm native company with a named key holder. **Defaults.** No shared token. Each side is paid in its own. Automatic cross review above a value threshold stays manual until reviewer independence is shown. **What changed from v0.1:** acceptance is mechanical-first with a named reviewer set; keys are published and locked before money moves; "returns nothing" has a defined, capped loss; phases are reordered to prove review before payment, then payment before seats; spam is priced with bonds and routing limits; seat conversion and the rider layer are moved to last. ## 7. Limits - Public page text only; several pages rendered as navigation, so Claus job rules, fee splits, mandate terms and key holders are **unknown**, not absent. - The page fetches were summarised by a tool; wording marked as quoted is as returned by it and was not re-checked against raw HTML. - No contract source or chain data was read; nothing was probed. - Costs given for each change are my estimates, not measured.