# Claus as a swarm beside IMD: how it breaks, and what to cut Final report. It brings together five separate views on DESIGN v0.1 (ultrapara), all written with the adversary-and-failure lens. Date: 12 November 2026. ## How to read this report - **[R]** means a cited public source says it, or a researcher's read-only chain call returned it. **[I]** means an inference. **[P]** means a proposal. **[?]** means it is unverified or the sources conflict. - **V1 to V5** are the five views (`artifacts/view-1.md` to `view-5.md`). Each researcher worked alone. Every claim carries the views that make it and the sources they cite. I did not repeat the views' chain reads. When two or more views report the same address from separate reads, I treat that as corroborated. - **My own check.** On 12 November 2026 I made one public GET each of `claus.si/migration-reference.json`, `claus.si/llms.txt` and `claus.si/Hooks`. The only purpose was to settle the fee-split conflict in correction C3. Nothing else was fetched, and I logged into nothing. - **Limits.** This is not an audit. None of the five views tested a paid delivery, measured common ownership between the two holder sets, or reproduced contract builds. V4 could not render the Claus Jobs or Hooks pages and read no chain data, so where it conflicts with views that read sources directly, I weight it lower. The rule for ranking changes: strength of evidence first, then the number of views. --- ## 0. Wrong or stale facts in v0.1, corrected first | # | v0.1 says | Correction | Views | Sources | |---|---|---|---|---| | C1 | Claus has "no independent acceptance or onchain record" | **[R]** Claus Jobs already has these: a reviewer committed in `reviewPolicy` before the claim; frozen input and output hashes; immutable, per-criterion reports; a rule that the reviewer cannot be the requester, the worker or a known controller; separately paid review commissions; and onchain delivery commitments for paid tasks. Reviews are **advisory** and only the requester accepts. **[I]** The real gap is *binding* independent acceptance plus a record readable across swarms. Reuse the existing primitive. | V1, V2, V5 (V3 in part) | [Jobs](https://claus.si/docs/jobs) | | C2 | No bridge exists yet | **[R]** Claus paid-deal disputes already buy an IMD oracle panel (5 members, 4 must agree), and Claus then approves the split. **[I]** This is a narrow arbitration dependency. It does not show general two-way hiring, and a Claus judgment routed through IMD is not an independent second-swarm check on IMD. | V1, V5 | [Paid deals](https://claus.si/docs/paid-deals) | | C3 | 2% pool fee: half buy-and-burn, half compute and hosting | **[?] Claus's own sources conflict today.** I checked on 12 Nov. `migration-reference.json` says "1% buyback and burn plus 1% operating treasury", `liquidityShareBps: 0`. `/Hooks` says 1% burn, and 1% "accrues in ETH for CLAUS DEV to fund inference, compute, hosting and ongoing work". `llms.txt` says the Robinhood 2% is "all allocated to liquidity … No current treasury share". The legacy Ethereum pool charges **2.3%**, all of it as ETH to `0xBb14…F1C5`. **[I]** v0.1 matches the Hooks page, but the "compute and hosting" half is a treasury allocation, not a commitment. Any split can change at the will of a single key (C10). **What would settle it:** a reading of the keeper's actual transfers on Robinhood, plus one machine-readable fee state published by Claus. | V1, V2, V3, V5 (they disagree, see §0.1) | [migration-reference](https://claus.si/migration-reference.json), [llms.txt](https://claus.si/llms.txt), [Hooks](https://claus.si/Hooks) | | C4 | The 300 houses can become ERC-8004 seats | **[R]** The houses are ERC-721 NFTs on **Ethereum** (`0x8039…95d9`). Each is backed by 50,000 CLAUS and can be burned to redeem them. A sale does not carry the house's private projects or keys to the buyer. **[R/C]** V2 counted 185 staked houses and 147 distinct holder or beneficiary wallets at Ethereum block 26397881. The largest holds 22, and the top five hold 80. | V2, V3 (V5 in part) | [nft-config](https://claus.si/nft-config.json), [nfts.json](https://claus.si/nfts.json), [llms.txt](https://claus.si/llms.txt) | | C5 | Sponsored inference of up to $5 a day is enough for a seat | **[R]** The $5 per NFT per day cap is correct. Sponsorship needs a funded period and automatic top-up is off. The live profile is `claude-haiku-5.5` only, with at most 1,024 output tokens, 6 requests a minute, 1 at a time, and `supports300Independent: false`. IMD contract work requires a frontier model at medium or high effort, plus Foundry. **[I]** A sponsored house cannot do IMD-grade upgrade review. | V2, V3, V1, V5 | [NFT capabilities](https://claus.si/__network/v1/nft/capabilities), [NFT guide](https://claus.si/docs/nft-inference), [IMD docs](https://imd.fun/docs/llms.txt) | | C6 | Inference submissions switched off on 21 Oct | **[R]** This applies to *project* inference only. NFT sponsorship is a separate service. **[?]** V5 found that the public capabilities endpoint reports `inference.enabled: true`, which contradicts the status page. V2 did not confirm the date. **[P]** Budget the pilot with no sponsored inference. | V1, V2, V3, V5 | [inference status](https://claus.si/docs/inference), [capabilities](https://claus.si/__network/v1/public/capabilities) | | C7 | Claus x402 services are in USDG and there is "no posted price" | **[R]** The current service page offers Contract Control in **USDC on Base**, including an x402 route, at 250,000 micro-USDC (0.25 USDC) per unit. **[I]** The blanket USDG claim is stale. There is still no posted price for *upgrade review*. | V1, V5 | [Customer services](https://claus.si/docs/customer-services), [products](https://claus.si/__network/v1/public/commerce/products) | | C8 | "395 API operations" | **[R]** The release page of 26 Oct says 351. The live OpenAPI 0.41.0 contains 426 operations by both V2's and V5's count. **[I]** No count says anything about swarm capacity. | V2, V5 (V3 unverified) | [release status](https://claus.si/docs/release-status), [OpenAPI](https://claus.si/__network/openapi.json) | | C9 | Jobs escrow "ETH or CLAUS" (implied: on Robinhood) | **[R]** Every escrow example uses `chainId: 1`. The CLAUS escrow terms point to the **legacy Ethereum token** `0x1b54…`, not to the Robinhood CLAUS. The ETH escrow is `0x354f…9816`, its arbiter is `0xb104…e105` and its admission controller is `0xbb14…f1c5`. The terms mark the operational patch as "not independently re-audited". **[I]** Any Robinhood-side bridge crosses chains on the Claus side. | V2, V3, V5 | [Jobs](https://claus.si/docs/jobs), [Ethereum deal terms](https://claus.si/__network/v1/public/chains/1/deals), [escrow source](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/contracts/src/ClausDealEscrow.sol) | | C10 | (implied) Claus contracts are a stable base for prices | **[R/C]** The Robinhood CLAUS token `0x1630…EfFAC` is an EIP-1967 proxy. The token and its ProxyAdmin `0xcbc4…22a8` are owned by the `ClausAdministrator` contract `0x979d…dc46`, which is owned by `0xbb14…f1c5`. That address has no code on Robinhood. The hook ProxyAdmin `0xf2ea…4ac0` has the same owner (V5). An **operator** `0x8aca…4cfa` may also execute calls on managed targets, with no timelock and no review check (V5, from published source). The docs say the owner can upgrade and mint with no cap, and that the 1-hour notice is "not enforced". The *Ethereum* token's ProxyAdmin is renounced (hook-health at block 26,397,884). The older Ethereum statement on the paid-deals page is superseded. | V1, V2, V3, V5 (chain reads: V1, V2, V5) | [llms.txt](https://claus.si/llms.txt), [migration article](https://claus.si/migration-article), [administrator source](https://github.com/contractclaus/claus-claus-buyback-v52-review-97934309/blob/ac63a28c06970294e222de51368a233d0fd11572/baseline/ClausAdministrator.sol), [hook-health](https://claus.si/hook-health.json) | | C11 | IMD intake "on Ethereum and Robinhood Chain" (no owner mentioned) | **[R/C]** Intake v2 is at `0xa43e…de82` on both chains. The same values came back on both: owner `0x047f…54b7` (an EOA, per V2), writer `0x3d3c…25cb`, quote signer `0x169b…279f`, payee `0x4e0f…dbc`. The verified Ethereum source uses `Ownable2Step`. The owner can set the price, payee, writer, quote signer and callback gas. Only the writer calls `complete`. Payment is forwarded in the same transaction. The contract is not a proxy. V3 and V4 reported the owner as "unpublished". That is superseded by the chain reads, but the docs still do not state it. **[?]** Nobody verified that the Robinhood bytecode matches. | V1, V2, V5 (V3, V4: unknown) | [Intake source](https://etherscan.io/address/0xa43e6f75ee006411f79ac1c84120606c2330de82#code), [IMD docs](https://imd.fun/docs/llms.txt) | | C12 | Onchain ERC-8004 record "on both chains" | **[R]** IMD's ERC-8004 adapter, WorkRegistry and ReputationRegistry are listed on **Ethereum and Sepolia only**. One wallet writes every review: WorkRegistry writer `0x5da3…6d44`. Claus has no ERC-8004 record. ERC-8004 is a **Draft**. Its text says Sybil attacks are possible and that `getSummary` needs a client-address filter. | V2, V3 (ERC-8004: all five) | [IMD docs](https://imd.fun/docs/llms.txt), [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004) | | C13 | IMD "flat 0.5 IMD a request" with fee flows unstated | **[R]** The price is correct. The IMD docs contradict themselves: the Intake forwards "whole to the treasury" (payee `0x4e0f…`), yet the docs also say "Nobody takes a fee. There is no treasury." | V2, V3, V4, V5 | [IMD docs](https://imd.fun/docs/llms.txt), [IMD skill](https://imd.fun/skill.md) | | C14 | Seats are "on the holder's own machine" (implied: one seat, one operator) | **[R]** The NFTs were minted free, one per wallet, in May 2026, and they can be transferred. Pairing refuses a device that already serves another seat. A public community repo runs "several NFTs on one server". IMD routing is by tools, load, failures and turn, and "past wins count for nothing". **[I]** A seat is a wallet, not a person. Rider "priority" would be a new incentive to farm. | V2, V3, V4, V1 | [IMD docs](https://imd.fun/docs/), [worker.md](https://imd.fun/skill/worker.md), [worker repo](https://github.com/Identity-md/worker) | | C15 | Matt Huang's post of 12 Nov 2026 | **[?] The views disagree.** V1, V2 and V3 could not find it. V5 found X post `2120209025099043286` through X's public oEmbed and dates it **10 Nov 2026**. The embed shows only the opening thesis. **[I]** V5's citation is specific and checkable, so the date in v0.1 is probably wrong. Even so, the thesis is motivation, not evidence. **Settles it:** the proposer confirms the link. | V1, V2, V3 vs V5 | [post](https://x.com/matthuang/status/2120209025099043286), [oEmbed](https://publish.twitter.com/oembed?url=https%3A%2F%2Ftwitter.com%2Fmatthuang%2Fstatus%2F2120209025099043286&omit_script=true) | ### 0.1 Where the views disagree on facts - **Fee split (C3).** V1 (migration guide) and V5 (Hooks) say 1% burn and 1% treasury. V2 (llms.txt) says all of it goes to liquidity. V3 reports both. My own check found that all three Claus pages are live today and still disagree. Neither side is wrong about its source. The disagreement is in Claus's own documents. - **Who controls IMD Intake (C11).** V3 and V4 say it is unpublished. V1, V2 and V5 read it from chain. Both are true: the address is readable, but the custody behind it is undocumented. - **Was the Ethereum CLAUS ever upgradeable (C10)?** V3 saw a conflict. V5's hook-health reading at a later block resolves it as renounced on Ethereum. That says nothing about Robinhood. --- ## 1. Five changes to v0.1, ranked The ranking is by strength of evidence, then by the number of views. All five changes are proposals **[P]**. The evidence behind each is labelled. ### Change 1. Publish and bound every rule-changing key before any money crosses **Views: 5 of 5** (V1 #1, V2 #1, V3 #2, V4 #2, V5 #1). **Evidence: strongest.** Three views made independent chain reads that agree (C10, C11), there is verified source code, and both projects document it. - **Why [I].** On Claus, one EOA (through `ClausAdministrator`) plus a delegated operator can upgrade the token and hook, mint with no cap, and change fees, with no enforced delay. A "review of every upgrade", a CLAUS price and a burn/seat/treasury split mean nothing if one key can bypass all three. On IMD, one owner EOA can change the price, the payee and the writer, and the writer alone decides which result or failure a payer sees. V5 adds that the administrator's execute path has no review hook, so on Claus an upgrade review cannot be a veto today. - **What to do [P].** Each swarm publishes a signed, versioned **authority manifest** covering: chain, contracts, implementation and code hashes, owner/pending owner/operator/writer/quote signer/payee, what each role can do, multisig threshold, timelock, emergency powers, and the website, scheduler and moderation powers. Each pilot order pins one manifest version. A change to the manifest stops new orders. - Before any *automated* spending, credit, or execution that depends on a review: Claus moves the token, hook and ProxyAdmin owner and operator behind a multisig, adds an enforced delay, and caps the mint. IMD moves the Intake owner to a multisig and publishes rotation plans for the writer and quote signer. - **Disagreement: how long the delay should be.** V1, V2 and V4 propose at least 48 hours. V3 proposes at least 72 hours, because IMD results take "minutes to days" and a review must be able to finish inside the delay. **What settles it:** the measured 95th-percentile delivery time of IMD audit jobs during the pilot. - **Disagreement: whether the lock is a gate on the *first* payment.** V2 and V4 say yes: no Phase 1 money until keys are locked. V1 and V5 say publication is enough for a tiny, capped, manually approved experiment, and locking gates only automation. V3 puts the gate between Phase 1 and Phase 2. **My recommendation:** publish before the first payment and lock before automation. A capped manual purchase is a disclosed risk. Requiring a lock first would block the cheapest proof. - **Cost.** Publishing takes days. A multisig and timelock cost Claus fast upgrades, and Claus's journal says its hooks are meant to keep evolving (V3). IMD's cost is one `transferOwnership` plus documentation. ### Change 2. Name who loses money on failure; do not sell "one payment" yet **Views: 5 of 5** (V1 #3, V2 #5, V3 #5, V4 #3, V5 #3). **Evidence: strong.** IMD's skill and onchain docs, the verified Intake source, the escrow source, and the x402 docs. - **Facts [R].** IMD: "no refunds; a refused body is a spent payment". Status 1 is refusal and status 2 is no result. Failure callbacks are unsigned, and anyone can trigger one with a bad body. Results can be late or missing. `check` and `quote` are free, and a quote is valid for 10 minutes ([skill](https://imd.fun/skill.md), [onchain](https://imd.fun/skill/onchain.md)). Claus escrow: delivery is due within 30 days, and `settleExpired` refunds after that. After delivery the buyer has 24 hours, and **silence pays the provider**. A dispute has 3 days, and an unresolved dispute refunds the buyer. Settlement creates withdrawal credits, so funds are not pushed (V3, V5; [paid deals](https://claus.si/docs/paid-deals), [escrow source](https://github.com/contractclaus/claus-network/blob/b1852dda3878a8b7c4ecf4ab8e3d41afa6147b9f/contracts/src/ClausDealEscrow.sol)). Free Claus tasks move no money. x402 has no refund step ([x402](https://docs.x402.org/introduction)). V5 adds that dispute admission closes 3 hours before the deadline and processing stops 1 hour before, which leaves at most 2 hours, far less than IMD's delivery time. - **What to do [P].** Use an itemized order with separate lines for: the non-refundable IMD attempt fee, the escrowed Claus work, the review fee, the deadline, the acceptance window and the retry budget. Run `check` and `quote` before every IMD payment. **Never move value on an IMD failure callback.** Do not make a paid call downstream unless the step upstream returned an accepted artifact. Track separate states (paid, admitted, delivered, checked, accepted, credited, withdrawn, failed, timed out, disputed). Reconcile an uncertain payment by its original ID; do not retry it automatically. Assign a funded monitor for the 24-hour and 3-day windows. - **Disagreement: who carries the IMD loss.** - V1 and V5: a pilot sponsor pays it directly and absorbs it. One payment is deferred until a merchant of record funds a refund reserve. - V3 and V4: the rider's one payment goes into Claus escrow now. A coordinator pays the 0.5 IMD from a float and recovers it only on acceptance (V3), or only the capped per-request fee leaves the escrow (V4). - V2: use each direction's existing rails, and ask IMD for a free `job.continue` credit when a job ends `blocked`. - **What settles it:** whether a founder will fund a float and a loss ceiling (open question Q4). Until then, follow V1 and V5. - **Cost.** More payment steps for the customer, a small loss budget, and a person or agent watching the deadlines. ### Change 3. Do not convert the 300 houses; start with a small roster of operators on their own frontier plans **Views: 5 of 5** (V2 #2, V5 #2, V3 #4, and V1 and V4 under "cut"). **Evidence: strong.** A live capabilities endpoint (C5), the NFT configuration and a chain holder count (C4), and v0.1's own gap list. - **Why [I].** A house is an entitlement on Ethereum that can be redeemed for cash. It is not working capacity. Sponsored inference is Haiku-class and not built to run 300 at once. Holding is concentrated (top five: 80 of 300), so a 30-seat pilot drawn from holders could fill a review panel from a few wallets. Conversion is also the largest piece of new code and the largest Sybil surface. - **What to do [P].** Recruit 3 to 10 named operators on their own models, across at least 3 disclosed control domains (V1, V5) or at least 5 distinct beneficiary wallets (V2). Participation does not depend on owning a house. Houses keep their current role. - **Disagreement on scope.** V3 would still have Claus seats do *low-stakes* work (hosting checks, site review) in Phase 2, running seat software on the holder's own plan. V1, V4 and V5 drop house-to-seat mapping until after the proof. - **Cost.** Political: holders lose a perk they expected (V2). There is also some manual onboarding. ### Change 4. Define independence by controller, not by seat or swarm; make reviews mechanical, sealed and assigned **Views: 5 of 5** (V1 #2, V2 #3, V3 #3, V4 #1, V5 #2). **Evidence: moderate.** The weakness is documented: Claus's Jobs docs say "Separate identities do not prove separate real-world operators", ERC-8004 states its Sybil limit, IMD prevents self-review only per seat, and Claus lets the requester choose the reviewer ([Jobs](https://claus.si/docs/jobs), [agent economy](https://claus.si/docs/agent-economy), [ERC-8004](https://eips.ethereum.org/EIPS/eip-8004)). The defences are proposals that nobody has tested. How each attack works and its cheapest defence. All defences are **[P]**; the views are in brackets. | Attack | How it defeats v0.1 [I] | Cheapest defence | Residual risk | |---|---|---|---| | Fake or duplicate seats | One operator runs many NFTs or devices and appears as a quorum | Count one vote per **control domain**, not per NFT. Check current ownership at assignment (all 5). Treat a wallet that funded the seat within N days as the same operator (V3, V4) | Hidden common ownership survives this | | One owner in both swarms | Builds on IMD and approves on Claus. v0.1's "second review from the other swarm" fails without anyone noticing | Take the intersection of the public IMD holder list (`ownerOf` on `0x0000ec93…`) with the Claus beneficiary list at assignment time (V2). Add signed cross-holding declarations; an undeclared link found later voids that reviewer's past reviews (V3). Disclose overlaps rather than ban them (V4) | Wallets cost nothing, so a determined owner can evade this | | Reviewer collusion | A requester-chosen reviewer certifies anything. A cartel trades approvals | The requester does not pick the reviewer. Pay a fixed fee whatever the verdict (V1, V3, V5). Count only findings that come with a reproduction (V1, V2, V3, V4). Have a named tie-breaker. Avoid repeat reciprocal pairs | A small roster can still be captured | | Copied reviews | The second reviewer paraphrases the first, which is public | Commit-reveal: a salted hash bound to the job and artifact hashes, revealed only after every reviewer has committed (all 5). Credit the earliest commitment when findings are identical (V3) | It does not stop collusion arranged in advance | | Testing whether reviewers work | Silence passes | Seed known-bad canaries, about 1 in 10 jobs (V2), or planted defects in fixtures (V1, V5) | Costs about 10% more jobs | - **Disagreement: who chooses the reviewer.** - V2: random draw from the eligible set. - V1, V4, V5: a short published allowlist or roster, accepting centralisation at first. - V3: the *other* swarm's coordinator chooses. - **What settles it:** whether either founder will accept the other side choosing its reviewers (Q7). Until then, a published roster with a random draw within it fits V1, V2, V4 and V5. - **Disagreement: how to pay reviewers.** V1, V3 and V5 pay a fixed fee whatever the verdict, so findings cannot be bought or extorted. V4 pays only when a later independent signal agrees. V1 explicitly warns against paying per finding. **My recommendation:** a fixed fee, plus loss of standing on a missed canary, which is V2's version. - **Mechanical first (V4, with V2 agreeing).** For upgrade review, the main check is a reproducible build plus diffs of the bytecode, storage layout and authority, emitted as a hash-pinned report. Opinion comes second. V1 and V5 add that IMD's text verification checks only structure ([IMD reviews](https://imd.fun/docs/api/#reviews)), so an "accepted" report is not a correct one. - **Cost.** Small scripts (holder intersection, commit-reveal), a canary budget, maintaining the roster, and more latency. ### Change 5. Reorder: one bounded handoff on existing rails, with no new code, before seats, bridge or rider layer **Views: 5 of 5** (V1 #5, V3 #1, V4 #4, V5 #5, V2 in §1g and §3). **Evidence: moderate.** It rests on inference from C1, C2, C5 and C9 and on v0.1's own gap list. - **What to do [P].** Run a small Phase 1 that uses only the IMD Intake and Claus Jobs/escrow or the City publisher. Include a defect-recovery test, a conflict-rejection test and a non-delivery or reconciliation test, not just successes (V1, V5). Hold the card and evidence log *before* the first order, and test the card on a third provider early (V1, V5). - **Disagreement: what the first job is, and its direction. This is the main open conflict.** - **V3 (backed by V2):** IMD audits a real pending Claus token or hook upgrade, ordered over the Robinhood Intake, and Claus holds the upgrade until the report is public. The reason: v0.1 has Claus review upgrades its own single admin controls, which is self-review in its plainest form, while IMD already sells an `audit` template with `repoUrl`/`baseCommit`. - **V4:** Claus reviews one real IMD-built upgrade, by hand. - **V1, V5:** use a frozen, non-production or historical candidate with seeded defects. IMD builds, Claus publishes and a separately controlled checker reviews. Nothing live is approved. - **My recommendation [P]:** V3's direction on a frozen candidate. IMD audits the next Claus hook or token upgrade before deployment, and the run includes one seeded fixture (V1, V5). This removes the self-review problem, uses an existing IMD product with no new code, and points review at the highest-risk keys. It is advisory: no callback can execute anything. - **What would settle it:** whether Claus will hold an upgrade for a report (Q2). - **Disagreement: which chain.** V2 and V3 want the Robinhood Intake. V5 wants Ethereum (IMD Intake plus an ETH Jobs agreement) because the Claus escrow exists only on Ethereum and CLAUS escrow uses the retired token (C9). **My recommendation:** IMD orders can go on either chain. Claus-side escrow must be on Ethereum in ETH until a Robinhood escrow exists. - **Cost.** A smaller headline: no seats, no token routing. One IMD fee, operator time and fixtures. V1 and V5 also add a *demand* test: one real, independent customer buys, and is then watched to see whether they reorder. --- ## 2. What I would cut | Item in v0.1 | Decision [P] | Views | |---|---|---| | Piece 1: all 300 houses become seats | **Cut** from the first proof (Change 3) | 5 of 5 | | Piece 4: one CLAUS price over x402 with a burn/seat/treasury split | **Cut** for now. Use a per-service quote in an asset that works today (USDC, ETH, IMD). Bring back a CLAUS SKU only after Change 1 and once C3 is resolved. Fixed prices need limits on scope, size and revisions (V5) | V1, V2, V3, V4, V5 | | Piece 6: rider "priority, credit and a fee share" | **Cut** in the pilot. Keep a read-only, signed evidence ledger with one canonical job ID, children for retries, related-party flags and failures visible | 5 of 5 | | Piece 6: "one payment" | **Defer** (Change 2). V3 and V4 would keep it now through Claus escrow | V1, V2, V5 defer; V3, V4 keep in modified form | | Piece 5: the two-way bridge, automated, with paid callbacks | **Split.** Phase 1 runs one way, by hand. V3 notes there is no Claus x402 hosting or review service for IMD to call yet | V1, V2, V3, V4, V5 | | "Work above a set value gets its second review from the other swarm" as an automatic rule | **Hold** until controller independence works. Replace the value threshold with **risk classes**: a cheap review of a mint key covers enormous risk, and orders can be split to stay under a threshold (V1, V5) | 5 of 5 hold. V3 keeps it as the right *eventual* rule | | Phase 2: 30 seats | **Replace** with 3 to 10 roster operators | 5 of 5 | | Phase 4: "first swarm native company … no staff" | **Cut** the "no staff" claim. Nobody can say who holds a company's upgrade key, handles incidents or owns it legally (V4, V5) | V1, V2, V4, V5 (V3 defers it) | | Phase 4: decision on a second seat class | **Cut.** Nothing earlier produces evidence for it | V1, V3 | | Free board tasks feeding swarm jobs | **Never.** Only paid or escrowed tasks from an admitted requester enter the swarm queue | 5 of 5 | **The open City: where spam gets in [R/I].** Wallet login needs only a signature from any EOA. Free tasks need no funded wallet. The board limits are 200 changes per participant per day, 1,000 a day board-wide, 5,000 stored records and 4,000-character briefs, and a claim locks a task for 1 hour ([Jobs](https://claus.si/docs/jobs), [build.txt](https://claus.si/build.txt)). The inferred attacks: - about five wallets exhaust the daily budget (V2); - claim-and-sit blocks tasks for free (V3, V4); - self-posted and self-accepted portfolios (V2, V3); - prompt injection through briefs that end up in IMD builder prompts (V2, V3, V5); - recursive "hire A to hire B" spending (V1, V5). **Defence [P]:** keep discovery open, and put a separate, initially manual gate before the paid queue. Caps apply per operator, not per wallet. Review agents get read-only access and no wallet. **Disagreement:** V4 wants a refundable claim bond now. V5 would add deposits only if spam is actually measured. V3 would require holding a seat to claim. Cheapest first: V5's manual gate, which needs no code. ## 3. What I would keep exactly as it is | Keep | Views | |---|---| | No shared token, so one side's key failure cannot spread to the other | 5 of 5 | | The swarm card's six questions, word for word. Add the authority manifest as a linked extension (V1, V4, V5) or as a seventh field (V3; V2 adds it under "who checked it"). Recommendation: an extension, so the six questions stay as they are | 5 of 5 | | Phase 0 as research ending in an explicit yes or no | 5 of 5 | | Upgrade review as the one first job type, advisory, on a frozen candidate | V1, V2, V4, V5 (V3 keeps it but reverses the direction) | | The division of labour: IMD builds and audits, Claus houses, operates and publishes | V2, V3, V4, V5 | | "Still live later" as part of the record. The checks: V2 at days 30 and 90, V3 at day 90, V5 at days 7 and 30. Bind each check to the original artifact hash and a functionality test, not just to whether a URL resolves (V1, V5) | V2, V3, V4, V5 | | A link between two swarms must work for a third | V1, V5 | | IMD's rule that "past wins count for nothing" in routing, which is a strength (V3) | V3 | **Disagreement:** "each swarm is paid in its own token". V2, V3 and V4 keep it. V1 and V5 relax it for the pilot, because the working rails are USDC on Base and ETH on Ethereum (C7, C9). **What settles it:** whether Claus ships a Robinhood CLAUS escrow (Q6). ## 4. Open questions only the founders can answer 1. **Claus:** Will the token, hook and ProxyAdmin owner (`0xbb14…` via `0x979d…`) and the operator `0x8aca…` move to a multisig with an enforced delay and a mint cap? When, and with which signers? (V1, V2, V3, V5) 2. **Claus:** Will an upgrade review become a technical prerequisite, or stay advisory? Who can bypass it? Will Claus hold its next upgrade for an IMD audit? (V1, V3, V5) 3. **Claus:** Which fee state is current: 1% burn and 1% treasury, or all to liquidity (C3)? (V2, V3, mine) 4. **Both:** Who pays the non-refundable IMD fee when a cross-swarm order fails, what is the maximum pilot loss, and will anyone fund a float or refund reserve? Will IMD give a `job.continue` credit for jobs that end `blocked`? (all 5) 5. **IMD:** How are the owner `0x047f…`, writer, quote signer, payee and WorkRegistry writer keys held? Will the owner move to a multisig? Is the Robinhood bytecode the same as on Ethereum? (V1, V2, V3, V4) 6. **IMD:** Is there a treasury or not (C13)? Where does the 0.5 IMD go? (V2) 7. **Claus:** Is there, or will there be, a Jobs escrow on Robinhood Chain that pays in the current CLAUS? (V2, V3) 8. **Both:** Who picks cross-swarm reviewers, and who can override the choice? Will each side accept the other choosing? Will you publish the overlap between IMD seat holders and Claus house beneficiaries once? (V2, V3, V4) 9. **Claus:** Will sponsorship stay Haiku at $5 a day? Do holders accept that reviewers run on their own plans and that being a reviewer is separate from owning a house? (V1, V2, V3) 10. **Both:** Will failed, timed-out, disputed and related-party jobs be published alongside the successes? (V1, V5) 11. **Both:** What does "rider owns it" mean legally, and who holds a swarm native company's keys and incident duty? (V1, V4, V5) 12. **Proposer:** Confirm the Huang post (C15). ## 5. Revised design, one page (v0.2) [P] *This is a proposal brought together from five views. Neither team has agreed to any of it.* **Goal.** Show that separately controlled operators can buy, check and publish one useful artifact across IMD and Claus, with bounded loss and attributable evidence, through an interface a third swarm can use unchanged. No autonomous company yet. **Before any money moves (Phase 0, ending in a yes or no).** - Correct C1 to C15. - Each swarm publishes the six-question **swarm card** plus a signed **authority manifest**: every privileged address, its powers, its multisig or timelock, and its emergency powers. Each pilot order pins a manifest version, and any change stops new orders. - Agree the rubric, the loss cap, the conflict rule, who bears the IMD attempt fee, and who can stop the pilot. **Phase 1: one handoff on existing rails, no new code.** - Claus buys IMD `audit` jobs (0.5 IMD each, after the free `check` and `quote`) on its next hook or token upgrade, frozen at a commit, plus one seeded-defect fixture. - The report is published in the City or the Claus journal. Claus holds the upgrade until the report is out. - A second check comes from a roster operator in a different, disclosed control domain, using commit-reveal and paid a fixed fee whatever the verdict. - Three IMD builds are published in the City. - Any Claus-paid leg uses the existing Ethereum ETH escrow, with a named monitor for the 24-hour and 3-day windows. **Pass** if all of the following hold: - the seeded defect is found, or the check honestly reports "inconclusive"; - a planted conflict is rejected; - receipts match the delivered bytes; - every attempt reaches an accounted end state, with no duplicate charge; - failures are visible. **Gate before automation.** Claus puts upgrade, mint and fee authority, including the operator path, behind a multisig with an enforced delay of at least 48 to 72 hours (set from measured IMD delivery time) and a mint cap. IMD moves the Intake owner to a multisig and publishes key custody. Without this, work stays manual and advisory. **Phase 2: small roster, both directions.** - 3 to 10 operators on their own frontier plans, at least 3 control domains, chosen at random from a published roster. - Cross-holding is checked by intersecting the two holder lists. - Every blocking finding needs a reproduction, and about 10% of jobs are canaries. - IMD buys Claus review or hosting through Claus escrow. - No paid callbacks, and no recursive orders. **Phase 3: record and third swarm.** - One canonical job ID, mirrored by both swarms as signed events: paid, admitted, delivered, checked, accepted, failed, disputed, still live at a set day. - Related-party work is flagged, and readers filter ERC-8004 summaries by published client addresses. - **No priority, credit or fee share.** - A third provider implements the card and receipt without any bilateral exceptions. - One real independent customer buys, and is watched to see whether they reorder. **Phase 4: separate decisions, each on its own evidence.** A CLAUS price and split, a single-payment merchant with a funded refund reserve, houses as a seat class, rewards from the record (counting only reviewed, unlinked, outside revenue, capped per counterparty), and any swarm-operated business with named human owners for keys, incidents and law. **Never.** No free board post becomes a swarm job. No IMD failure callback moves value. No review or callback executes an upgrade or releases treasury funds. No refund is promised without a funded guarantor.