# Claus as a swarm adjacent to IMD: what I would change in the economics Researcher 1 of 5, working alone. Lens: economics and incentives. Written 2026-11-12, from public pages read that day. **How to read the labels.** Every statement carries one of these tags: - **[R: Sx]**: read from source Sx (listed at the end), as published on 2026-11-12. - **[I]**: my inference from the sources. It may be wrong. - **[P]**: my own proposal. - **[U]**: unverified or uncertain. I looked and could not confirm it. **Limits.** This is desk research on public pages only. I did not log in, connect a wallet, call a paid or write endpoint, or read chain state. I did not check token prices, so the report has no USD value for 0.5 IMD or for any CLAUS amount, and every number I propose is a starting point to calibrate, not a result. The "Matt Huang, Nov 12 2026" post that the design builds on did not turn up in a web search (S15), so I treat its thesis as the design author's premise and not as a cited fact **[U]**. --- ## 0. Corrections first: where v0.1 rests on wrong or stale facts | # | v0.1 says | What the public sources say | Why it matters for the economics | |---|---|---|---| | C1 | Claus has "a 2% pool fee, half to buy and burn CLAUS, half to compute and hosting". | Robinhood official pool: "2% project fee, all allocated to liquidity… No current treasury share" **[R: S1]**. The Ethereum pool charges 2.3%, all of it accruing as ETH to one fixed recipient (CLAUS DEV). New NFT, LP reward and buyback allocations are paused **[R: S1]**. Claus's stated priority is "build Robinhood liquidity with the complete 2% project fee" **[R: S1]**. | Today no pool fee funds burns or compute on Robinhood. Any split in v0.1 that leans on "the burn half" or on treasury-funded compute rests on a mechanism that is not running **[I]**. | | C2 | "The 300 houses become seats under ERC-8004." This implies Robinhood-native NFTs. | The 300-NFT collection is deployed on **Ethereum** (chainId 1, capacity 300). Each NFT is backed by 50,000 of the Ethereum CLAUS token, and the market fee is 2% **[R: S2]**. "NFTs, backing and stakes remain on Ethereum" **[R: S1]**. Each NFT represents 50,000 deposited CLAUS and can be burned to redeem them **[R: S1]**. | A seat therefore carries a floor value (50,000 CLAUS of backing) and can be redeemed out of existence. This shapes seat scarcity and how to design a second class (§1, change 3) **[I]**. | | C3 | "Sponsored inference up to 5 dollars a day" is treated as a seat's compute. | Up to $5 per NFT per day, default model `claude-haiku-5.5`, max 1,024 output tokens, 6 requests a minute, concurrency 1 **[R: S3, S4]**. "Simultaneous 300-request execution is not confirmed" (`supports300Independent: false`) **[R: S4, S5]**. "Sponsored inference does not host an always-running agent". Only an explicitly funded period is available, and "a review date does not promise renewal" **[R: S6]**. | The sponsored allowance cannot run a build or review worker of the class IMD requires. IMD lists Fable 5.1 high, Opus 5.5 medium and GPT-6 class models **[R: S8]**. So the "treasury pays" option is not just costly; it is the wrong tool **[I]**. | | C4 | "New project inference submissions switched off on Oct 21 2026." | Confirmed: "New project inference requests are disabled as of 21 October 2026" **[R: S7]**. The public capabilities JSON still reports an `inference.enabled: true` flag under a project-allowance profile **[R: S16]**, and the product reference says `projectInference.status: "enabled"` **[R: S5]**. | The disable notice is the clearer source. The machine-readable flags disagree with it, and a coordinator must not route on them **[I/U]**. | | C5 | IMD "flat 0.5 IMD a request" is the model for a split "between burn, seats and treasury". | IMD's onchain intake forwards the 0.5 IMD "whole to the treasury in the same call" **[R: S8]**. The same docs also say "Nobody takes a fee. There is no treasury" **[R: S8]**. Publicly, seats earn **reputation** plus **8% of each launched token** (seats connected at admission) and **2%** (wallets with accepted work on that launch) **[R: S8]**. No per-request cash payout to seats is published, and the worker is "paid by the network" with "no guaranteed earnings" **[R: S9]**. | IMD does not split request revenue among burn, seats and treasury in any published way. Its seat income depends on launches. v0.1 should not present its own split as "like IMD", and IMD's two contradictory treasury statements need an answer from the founders (§4) **[I]**. | | C6 | Claus has "no independent acceptance". | Claus Jobs lets a requester commit a `reviewerId` and criteria in a `reviewPolicy` before a claim. Reviewers can be matched by role, and review work is paid "regardless of its conclusion". Reports, however, "are advisory; they cannot sign for the buyer" **[R: S10]**. Only the requester can accept **[R: S11]**. | The gap is narrower than v0.1 says. Claus already has reviewer roles and paid review. What it lacks is **binding** acceptance and a seat-level onchain score **[I]**. | | C7 | The bridge and escrow are on Robinhood Chain. | The Claus Jobs escrow examples and `deal_terms` all use `chainId: 1` (Ethereum), and gas is in ETH **[R: S11]**. IMD's intake is sold on Ethereum and Robinhood Chain **[R: S8]**. | Paying across swarms in CLAUS-on-Robinhood may not match where Claus escrow lives today **[U]**. This needs checking before phase 1. | | C8 | (not in v0.1) | The Robinhood CLAUS token, hook and liquidity management are owner-upgradeable by "a single admin wallet with extra mint authority and no onchain mint cap or timelock". The one-hour notice is "not enforced by the contracts" **[R: S1]**. | This is the main counterparty risk for any IMD balance held in CLAUS, and it bears directly on "can one side drain the other" (§1, change 2) **[I]**. It is also why upgrade review is a good first job. | --- ## Answers to the lens questions (short) **Are 300 seats enough, and what second seat class?** - 300 is enough for **one** job type such as upgrade review, but not for a general swarm **[I]**. - IMD refuses an oracle quorum above "the seats that took work in the last 24 hours" **[R: S8]**. This suggests the binding number is *active* seats, not issued seats **[I]**. - If 20 to 40% of 300 holders run a worker, which is a guess I cannot check **[U]**, that gives 60 to 120 active seats. That covers 3 to 5 reviewers per upgrade review at low volume, but not builds plus cross-review plus a rider layer **[I]**. - Best second class: **delegated seats** (a holder leases a seat's work right to an operator) plus **bonded probation workers**. Do not grant seats for burning CLAUS. See change 3 **[P]**. **What should a request cost and how should it split?** - Post a price per job type, fixed in USD and payable in CLAUS (or USDG) at a quoted rate **[P]**. - Split: **70% to the seats whose work was accepted** (builder and reviewers), **10% to CLAUS liquidity or burn** (follow whatever the pool currently funds, liquidity today per C1), **20% to the coordinator and treasury** **[P]**. - IMD by contrast sends 100% of the request fee to a treasury address and pays seats through launch allocations **[R: S8]**. **Who pays for inference, and would holders run a worker?** - The seat holder pays, on their own plan, as on IMD **[R: S8]** **[P]**. - Holders will run a worker only if expected per-job income beats the plan cost. Today a Claus holder has no posted per-job payout, so I see no reason they would **[I]**. - Keep the $5/day sponsorship as a trial and onboarding allowance, not as worker compute (C3) **[P]**. **How does value move without a shared token, and can one side drain the other?** - At the boundary, the buyer **buys the seller's token at market** (or pays a stable asset that the seller converts). Each swarm is paid only in its own unit **[P]**. - Drain paths exist **[I]**: - IMD's flat 0.5 IMD buys a large workflow at the same price as a small job **[R: S8]**, so a heavy buyer pays below cost. - Claus's admin can mint without a cap **[R: S1]**, so CLAUS held by IMD can be diluted. - Fix with per-type prices at the boundary, net settlement each epoch, a cap per epoch, and no CLAUS balances held by IMD beyond one epoch (change 2) **[P]**. **How should a rider record pay out?** - Priority and credit come from accepted and shipped jobs **[P]**. - Cash fee share comes only from outcomes still live and used at 30 and 90 days, capped, funded from the outcome's own revenue, and reduced by rejections **[P]**. - Volume alone earns nothing **[P]**. **What does a swarm native company owe each swarm, and how do launch allocations work across two swarms?** - Keep one swarm pool the size of IMD's 10% (8% connected seats + 2% accepted work) **[R: S8]**. Split it between the two swarms in proportion to accepted work receipts weighted by posted price **[P]**. - After launch, the company owes only posted-price fees for services it keeps buying (hosting, re-review). No perpetual royalty **[P]**. --- ## 1. Five changes to v0.1, ranked ### Change 1: Pay seats per accepted job; seats pay their own inference **[P]** **What.** - Each Claus job type has a posted price **[P]**. - When a job is accepted, 70% of the fee goes to the seats with accepted work on it, split by role: builder 50%, reviewers 20%, shared equally **[P]**. - Rejected work earns nothing, as in IMD's `acceptance-v2` scoring of 1 or 0 **[R: S8]**. - Inference runs on the holder's own plan. The $5/day sponsorship becomes a 14-day trial for new seats **[P]**. **Why.** - v0.1 asks holders to bring their own plan (piece 1) but never says what they are paid. Reputation alone did not create a worker market on Claus: there is "no working seat" (v0.1's own gap list) **[I]**. - IMD's seat income is lumpy and depends on launches (C5). Claus has fewer launches tied to its seats, so it needs cash per job **[I]**. - The sponsored route is Haiku-only, capped at 1,024 output tokens and not confirmed for 300 concurrent requests (C3). It cannot do upgrade review at quality **[I]**. **Cost.** - The treasury gives up most request revenue (it keeps 20%). - Someone must run a verifier that decides "accepted". Change 4's binding acceptance is a prerequisite. - Holders who will not pay for a plan drop out, which shrinks active seats. ### Change 2: Settle across swarms at the boundary in the seller's unit, netted and capped **[P]** **What.** - When Claus orders IMD work, it pays the IMD intake in IMD. Claus's router buys the IMD at market from a stable or CLAUS balance, at order time **[P]**. - When IMD orders a Claus service, it pays in CLAUS or USDG over x402 at Claus's posted price. x402 is network- and token-agnostic, so both are possible **[R: S13]**. - Reciprocal second reviews (v0.1 piece 5) are **netted per epoch**: count reviews each way and pay only the difference **[P]**. - Each side caps per-epoch spend on the other, e.g. a fixed USD amount set by each founder **[P]**. - IMD converts any CLAUS it receives within the epoch **[P]**. **Why.** - "No shared token" is right, but v0.1 never says how value crosses. Buying the seller's token at market is the only way that keeps each token's demand tied to its own swarm's work **[I]**. - Drain risk is real on both sides: - IMD's single 0.5 IMD price for any request, "whatever the panel size or number of steps" **[R: S9]**, lets a heavy buyer (Claus's coordinator sending big workflows) consume more seat work than it pays for **[I]**. - CLAUS has uncapped admin mint with no enforced notice **[R: S1]**, so CLAUS held by IMD is exposed to dilution **[I]**. - Netting cuts the token flow, and the slippage, of reciprocal reviews to near zero **[I]**. **Cost.** - Market buys at each boundary add slippage and gas. - Netting needs a shared ledger: a signed epoch statement both sides publish. - IMD would likely need per-type pricing for cross-swarm orders, which departs from its flat price. ### Change 3: The second seat class is delegated and bonded, not burned **[P]** **What, in two tiers.** - **Delegated seats.** Any of the 300 holders can lease the seat's work right to an operator for a term, e.g. 30 days. This mirrors Claus's existing 30-day mandate idea. Reading Claus mandates as 30-day comes from the brief; I did not find the term on the pages I read **[U]**. - The operator gets the per-job payout. The holder takes a set cut, e.g. 20% **[P]**. - Reputation accrues to the ERC-8004 agent of the operator, not to the NFT **[P]**. ERC-8004 identities are separate ERC-721 handles **[R: S12]**. - **Bonded probation workers.** Anyone without a seat can stake CLAUS as a bond. The bond is forfeited on rejected or abusive work, not burned on entry **[P]**. - They take only low-value work, at a lower payout share **[P]**. - After N accepted jobs they can lease a seat or buy one. They never mint a new seat **[P]**. **Why.** - Burn-to-seat turns a one-time sink into a permanent claim on future fees. It dilutes the 300 and caps nothing **[I]**. - The 300 are already backed by 50,000 CLAUS each and can be redeemed by burning (C2), so seat supply can only shrink **[R: S1, S2]** **[I]**. - Delegation raises *active* seats without raising *issued* seats, and active seats are what IMD's quorum rule counts **[R: S8]** **[I]**. **Cost.** - Needs a lease contract (or an ERC-8004 operator field), plus bond custody and forfeiture rules. - Holders' share is diluted only by their own choice to lease. - The forfeiture judge must be credible, which ties this back to change 4. ### Change 4: Binding acceptance with a seat-level score, before any money flows **[P]** **What.** - Claus adds a binding acceptance path for coordinated jobs: an independent second run plus K reviewers. The reviewers are committed before claim, as `reviewPolicy` already supports **[R: S10]**. - Pass or fail is written to the ERC-8004 Reputation Registry. ERC-8004's Reputation Registry is "a standard interface for posting and fetching feedback signals" **[R: S12]**. IMD's WorkRegistry receipt is a working pattern for this **[R: S8]**. - Payout and rider credit key off this verdict, not off the requester's click **[P]**. **Why.** - Today only the requester accepts, and reviewer reports "are advisory" **[R: S10, S11]**. A payout that depends on the buyer's acceptance invites the buyer to withhold acceptance and the worker to collude with the buyer **[I]**. - v0.1 has acceptance as piece 3, after price (4) and seats (1). It must come first, because changes 1, 3 and 5 all pay on it **[I]**. **Cost.** - The second run and the reviewers roughly double inference per job. That cost goes into the posted price **[I]**. - Writing to a reputation registry costs gas. IMD batches up to 64 reviews per receipt **[R: S8]**, and Claus should do the same **[P]**. ### Change 5: Rider record pays only on outcomes that last; one swarm pool at launch, split by work **[P]** **What.** - **Rider tiers (all [P]):** - Priority: queue order comes from accepted jobs in the last 90 days. - Credit: up to the value of the rider's accepted jobs in the last 30 days. - Fee share: a rebate of up to 5% of fees on the rider's jobs. It vests only when the shipped output is still live and used at 30 and 90 days, checked by an oracle question. IMD sells oracle questions **[R: S8]**. - Rejected or abandoned jobs subtract. - **Launch allocation for a swarm native company:** - One swarm pool of 10%, IMD's current size **[R: S8]**, not 10% from each swarm **[P]**. - The pool splits between IMD and Claus in proportion to price-weighted accepted receipts on that company **[P]**. - Within each swarm, the swarm's own rule applies. For IMD that is 8:2 connected-seats to accepted-work **[R: S8]**. For Claus I propose accepted-work-only, because "connected" is not a meaningful state for Claus houses yet **[P/I]**. - **After launch:** the company owes each swarm only posted-price fees for services it keeps buying, such as hosting, re-review or upgrade review. No royalty **[P]**. **Why.** - Rewards on volume are easy to game with self-ordered jobs **[I]**. Liveness at 30 and 90 days is a cheap proxy for value that is hard to fake **[I]**. - Stacking two swarms' allocations would take 20% of every joint company's supply. That makes the joint product worse than launching on one swarm **[I]**. **Cost.** - Rewards are delayed by 90 days. - Liveness oracles cost requests. - The rider gets less at launch than a naive "fee share" promises. - IMD would need its launch distributor to accept a second swarm's recipient list. --- ## 2. What I would cut **[P]** - **"One posted price per request in CLAUS."** Replace it with a price per job type, fixed in USD, paid in CLAUS or USDG. A CLAUS-denominated price swings with a token whose supply the admin can expand **[R: S1]** **[I]**. - **A split that names "burn".** Replace it with "liquidity or burn, whichever the pool currently funds". Today it funds liquidity (C1). - **"Work above a set value gets its second review from the other swarm" as a rule.** Make it an option the buyer prices in. A mandatory cross-review gives each swarm a guaranteed income stream from the other, regardless of quality **[I]**. - **"A rider record kept by both."** Keep one ERC-8004 record per rider per swarm with a shared schema, so a third swarm reads it without a joint database **[P]**. - **Phase 4's "decision on a second seat class".** Move it to phase 2. The pilot will show whether active seats run short, and leasing should already be possible by then **[P]**. - **Treasury-funded worker inference** (the implicit status quo). Keep sponsorship only as a trial (change 1). ## 3. What I would keep exactly as is **[P]** - **No shared token; each swarm is paid in its own.** - **The swarm card's six questions** (who are your seats, what do you sell, what does it cost, how do I pay, what do I get back, who checked it). They map cleanly onto what IMD already publishes: capabilities, price, intake, delivery, WorkRegistry **[R: S8]** **[I]**. - **Upgrade review as Claus's first coordinated job.** CLAUS's token, hook and liquidity are owner-upgradeable with no enforced timelock **[R: S1]**. Paid outside review of each upgrade is the most useful job Claus seats could do, and holders benefit directly **[I]**. - **Phase 0 (research, then yes or no) and phase 1 (hire each other with what exists).** Both swarms can already buy from each other: IMD's intake is on Robinhood Chain **[R: S8]**, and Claus Jobs and x402 services exist **[R: S11, S5]**. - **ERC-8004 as the seat identity.** It is already IMD's registry for seats **[R: S8]**, and it lets a third swarm read records without trusting either site **[R: S12]**. ## 4. Open questions only the founders can answer **IMD** 1. Does the 0.5 IMD "forwarded whole to the treasury" **[R: S8]** fund seats, burn, or operations? The same docs say "there is no treasury" **[R: S8]**. Which is current, and would IMD publish the use of request revenue? 2. Would IMD accept per-type pricing for cross-swarm orders, or a per-epoch cap on one buyer? 3. Would IMD's launch distributor accept a second swarm's Merkle list for joint companies? **Claus** 4. Will Claus put a cap or enforced timelock on admin mint and upgrades? If not, what would make CLAUS acceptable for IMD to hold? **[R: S1]** 5. Seats and backing are on Ethereum **[R: S1, S2]**. Will seat registration (ERC-8004) and seat payouts live on Ethereum or Robinhood? 6. How long will the $5/day NFT sponsorship stay funded, given "no renewal promise" **[R: S6]**? Can it be repurposed as a trial? 7. Is Claus escrow available on Robinhood Chain, or only on Ethereum as the examples show (C7)? 8. What USD prices per job type would each founder set? Who sets the per-epoch caps? 9. Will Claus holders accept a 70% seat share that leaves the treasury 20%? **Both** 10. Are both teams willing to let an independent verifier, rather than the buyer, decide acceptance and payout? --- ## 5. Revised design, one page **[P unless marked]** **Goal.** Claus becomes a second swarm beside IMD, specialised in operating and reviewing what is launched. Each swarm keeps its own token. The link is a published swarm card plus boundary settlement that a third swarm can copy. **Facts it rests on.** - Claus Robinhood pool: 2% to liquidity **[R: S1]**. - 300 NFT houses on Ethereum, each backed by 50,000 CLAUS **[R: S1, S2]**. - NFT sponsorship: Haiku-only, $5/day, not for workers **[R: S3–S6]**. - IMD: 0.5 IMD flat to a treasury address; seats paid by 8%+2% launch shares and reputation **[R: S8]**. **Order of work.** 1. **Acceptance first.** Binding verdicts on Claus coordinated jobs: independent second run plus K committed reviewers. Verdicts are written to the ERC-8004 reputation registry in batches. 2. **Price and split.** A price per job type, fixed in USD and paid in CLAUS or USDG over x402. Of each accepted job's fee: - 70% to the seats with accepted work (builder 50, reviewers 20) - 10% to CLAUS liquidity or burn, following the pool's current mode - 20% to the coordinator and treasury Rejected work earns nothing. IMD keeps its own model. 3. **Seats.** - The 300 houses register as ERC-8004 agents and run on the holder's own plan. - Delegated leases (holder cut, e.g. 20%) raise active seats. - Bonded probation workers (forfeitable CLAUS bond, low-value work only) are the entry path. - No new seats are minted or granted for burning. 4. **First job.** Paid review of every CLAUS token, hook or vault upgrade, plus upgrades for outside buyers. 5. **Cross-swarm settlement.** - The buyer pays the seller's posted price in the seller's unit, bought at market at order time. - Reciprocal reviews are netted per epoch. - Each side caps its per-epoch spend on the other and converts the other's token within the epoch. - Cross-review is offered and priced, never mandatory. 6. **Riders.** - Priority from accepted jobs (90 days). - Credit up to accepted value (30 days). - A rebate of up to 5% of fees, vesting only if the output is live and used at 30 and 90 days. - Rejections subtract; volume alone earns nothing. 7. **Swarm native company.** - One 10% swarm pool at launch, split between swarms by price-weighted accepted receipts. - Each swarm distributes its share by its own rule. - After launch, only posted-price service fees. The rider owns the rest. 8. **Swarm card.** The six questions unchanged, plus a machine-readable price list and settlement address per job type. **Phases.** - **0:** yes or no, and answers to §4 Q1, Q4, Q5. - **1:** three orders each way with what exists. - **2:** 30-seat upgrade-review pilot with binding acceptance and per-job payout. Leasing is switched on. - **3:** netted cross-review, rider record, and the card published. - **4:** all 300 seats plus delegates, the first joint company on the 10% pool, and the card opened to a third swarm. **Main risks.** - Too few holders run a worker at the posted payout. Watch active seats in phase 2. - CLAUS admin mint. Answer §4 Q4 before IMD holds CLAUS. - IMD's flat price is drained by heavy cross-swarm orders. Use caps and per-type pricing. --- ## Sources (all read 2026-11-12) - **S1**: Claus, `llms.txt` (migration, fees, admin powers, NFT backing): https://claus.si/llms.txt - **S2**: Claus NFT collection configuration: https://claus.si/nft-config.json - **S3**: Claus, NFT inference guide: https://claus.si/docs/nft-inference - **S4**: Claus, NFT capabilities (public JSON): https://claus.si/__network/v1/nft/capabilities - **S5**: Claus, product reference (public JSON): https://claus.si/product-reference.json - **S6**: Claus, NFT inference sponsorship conditions, version 2026-10-23-v1: https://claus.si/docs/nft-inference-terms - **S7**: Claus, project inference guide (disabled from 21 Oct 2026): https://claus.si/docs/inference - **S8**: IMD docs as Markdown (seat earnings, intake price, verifier, WorkRegistry, launch shares, contracts): https://imd.fun/docs/llms.txt, which mirrors https://imd.fun/docs and https://imd.fun/docs/api - **S9**: IMD skill file: https://imd.fun/skill.md - **S10**: Claus, agent economy (reviewer roles, `reviewPolicy`, advisory reports): https://claus.si/docs/agent-economy - **S11**: Claus, Jobs guide (claim lease, requester-only acceptance, ETH/CLAUS escrow): https://claus.si/docs/jobs - **S12**: ERC-8004: Trustless Agents (identity, reputation and validation registries): https://eips.ethereum.org/EIPS/eip-8004 - **S13**: x402 README (network- and token-agnostic, trust-minimising schemes): https://github.com/coinbase/x402 (raw README fetched from `raw.githubusercontent.com/coinbase/x402/main/README.md`) - **S14**: Claus home page (CLAUS on Robinhood Chain 4663, the old Ethereum token retired): https://claus.si - **S15**: Web search for the cited Matt Huang "swarm riders / swarm native companies" post returned no matching result. Nearest hits were unrelated profiles, e.g. https://www.paradigm.xyz/team/matthuang. Treated as unverified. - **S16**: Claus public capabilities JSON: https://claus.si/__network/v1/public/capabilities The brief also mentions 395 API operations, x402 services in USDG and 30-day mandates. I saw USDG x402 receipts only in a project listing in S5 and did not verify the other figures. I treat them as the brief's claims, not mine **[U]**.