# View 3: Claus/IMD swarm economics — independent review of v0.1 Researcher 3 of 5. Written alone, from public sources only. No private/write endpoints touched. ## Tagging convention Each claim is tagged: **[S]** read from a cited source, **[I]** my inference from sourced facts, **[P]** my own proposal/opinion, **[?]** unresolved/uncertain. ## 0. Sources actually retrieved and what they confirm - imd.fun landing page [S]: no economics on the page itself; links to `imd.fun/skill.md` and `imd.fun/docs/`. - `imd.fun/skill.md` [S]: price is flat **0.5 IMD per request** (0.5 IMD per question for `oracle.group`); price is independent of panel size/steps; live price served at `GET $IMD_API/requests/capabilities`; **no refunds** — "a refused body is a spent payment," payment "buys the work being admitted, not a guaranteed result"; worker pay exists but "a worker is not guaranteed earnings," no rate stated; intake-contract approvals are commonly capped at 5 IMD (10 requests); IMD token contracts given for Ethereum (`0xd34a99bc0f67ae1bbd63c660e6d0b0dd03e263b7`) and Robinhood Chain (`0x5f7bb59365ce557c26dbcaa4ee9d39a4b95b7127`), intake contract `0xa43e6f75ee006411f79ac1c84120606c2330de82`. Seat count and launch-allocation percentages are **not stated in this document** — it defers to `imd.fun/seats/` and to launch-specific pages I did not retrieve. ERC-8004 is not mentioned on this page. - claus.si landing page [S]: confirms CLAUS now lives on Robinhood Chain (chain id 4663) at `0x163015276f3244bC2CcB78b4D69cc8A4069EfFAC`; a per-trade pool fee funds buyback-and-burn plus "compute, hosting, and the work that keeps the agent running," but **no percentage split is given on this page** (the task brief's "2% pool fee, half burn / half compute" is asserted by the brief, not by what I could retrieve). - `claus.si/docs` [S]: landing/navigation only; house count, $5/day sponsorship, jobs-board escrow mechanics, API-operation count, mandate length, and the Oct 21 2026 cutoff were not present on the page retrieved, and sub-pages (Jobs, Hooks, Buyback-and-Burn, Migration) were not fetched due to time constraints. **Correction to v0.1 going in**: I could not independently verify several numbers v0.1 states as fact (2,000 IMD seats, 8%/2% launch split, 300 Claus houses, $5/day sponsorship cap, 2% pool fee with 50/50 burn/compute split, the Oct 21 2026 submissions cutoff, the "395 API operations"). I did verify the 0.5 IMD flat price, the no-refund policy, and that CLAUS is now single-chain on Robinhood Chain. **[?]** Treat the unverified figures in v0.1 as the requester's own representations until the founders confirm them; my recommendations below are written to be robust either way, but #2 and #5 below depend on the exact split numbers being right. ## 1. Answers to the LENS questions **Are 300 seats enough, and what's the best second class?** [P] 300 is enough to *start* a single job type (upgrade review) but not enough to be "a swarm" in the sense IMD's 2,000 seats imply — review needs redundancy (independent second run + reviewers per v0.1's own design), so 300 holders running real workers, after accounting for churn and non-participation, likely yields well under 100 live workers at any time. The better second class is not more seats but **delegated capacity under the same 300 seats**: let a house seat delegate its workload to a managed worker (operated by Claus City or a third-party operator) for a revenue share, keeping seat count and governance weight fixed while scaling throughput. Earning new seats by burning CLAUS **[P] rejected** — it dilutes the fixed 300 (same problem v0.1 is trying to avoid) and conflates "holds tokens" with "can run a reliable worker," which is exactly the scarce-labor problem Huang's post is about [S, as summarized in the brief]. **Price and split vs. IMD?** [P] Price the first Claus swarm job (upgrade review) as a **flat fee per review in CLAUS**, mirroring IMD's flat 0.5-IMD-per-request model (confirmed [S] in skill.md) rather than a variable/escrow price — flat pricing is what makes the "one posted price" promise in v0.1 operationally simple and auditable. Split: unlike IMD, which (per the brief, unverified) pays seats via launch allocations rather than per-request, Claus should split the posted fee into three parts paid out immediately per job, not accrued through a pool trade fee: burn / reviewing seats / treasury. I'd weight it toward seats (e.g., majority to the two reviewers who did the work) since the goal is to fund workers to run inference on their own machines, with smaller fixed shares to burn and treasury. **Who pays for inference?** [P] This is the single most important fix. v0.1 already flags the gap: "inference paid by the treasury instead of the seat holder." Treasury-sponsored inference (the $5/day cap, per the brief) is fine for a free/low-stakes jobs board but cannot survive contact with paid, competitive seat work — it caps throughput at whatever the treasury can subsidize and gives seat holders no reason to invest in better compute. Would Claus holders plausibly run a worker on their own plan? **[I]** Only if the per-job payout reliably exceeds their inference cost, which requires the posted price (above) to be real money flowing to the seat, not a token-burn mechanic. Treasury sponsorship should be phased out for seats that opt into paid review work, retained only for the free jobs-board tier. **Cross-swarm value flow without a shared token — can one side drain the other?** [I] Yes, structurally: if Claus seats disproportionately buy cheap IMD review/build jobs while IMD seats rarely buy Claus jobs (or vice versa), the paying side continuously exports value in its own token for work it can't reciprocate in kind. v0.1's mitigation — "work above a set value gets its second review from the other swarm" — only forces *reciprocal task flow*, not *reciprocal payment flow*; a swarm can still be a net payer indefinitely. **[P]** The fix is a rolling settlement cap or a quota (e.g., "no swarm pays the other more than N requests net per settlement window without the deficit being made up in kind or converted at a floor rate"), reviewed at each phase gate, rather than trusting volume parity to emerge on its own. **Rider record payouts?** [P] Pay out on *acceptance*, not on *claim* or *delivery* — v0.1 already has an acceptance step (independent second run + reviewers) so hook the record to that event. Priority and fee share should scale with a trailing acceptance rate (accepted/claimed over a rolling window) rather than raw accepted count, so a rider who ships 3 good jobs out of 3 outranks one who ships 30 out of 300. This directly targets "rewards shipped outcomes and not volume." **What does a swarm-native company owe each swarm, and launch allocations when two swarms contribute?** [P] It owes each swarm a continuing, metered share tied to usage it actually drives post-launch (hosting/review/build fees), not just a one-time launch allocation — a one-time grant gives both swarms incentive to over-launch and under-maintain. For launch allocations, split by verified contribution at launch time (IMD's build/audit work vs. Claus's housing/review work), using the seat/wallet mechanism IMD already has (8%/2% split per the brief, unverified) as the template, extended to credit both swarms' seats according to which one did the qualifying work on *that* company. ## 2. Five ranked changes to v0.1 **1. Make seat-funded inference the default for any paid job, not treasury sponsorship. (Highest priority)** Reason: this is the clearest economic inconsistency in v0.1's own gap analysis — a seat that doesn't bear its own compute cost isn't a real seat in the IMD sense, and a review system subsidized by the treasury can't price or scale honestly. Cost: requires Claus to define a seat-vs-sponsored-house split (e.g., sponsorship only for jobs-board/free tier; paid review requires self-funded inference), which may shrink the initial 300 down to however many houses are willing to self-fund — a real reduction in Phase-2 pilot capacity (v0.1 names 30 seats for the pilot, so this is survivable). **2. Replace the pool-trade-fee funding model for reviewer pay with a per-job posted price with an explicit split, paid at acceptance.** Reason: v0.1 already proposes "one posted price... with a stated split between burn, seats and treasury" (point 4) but the current Claus pool fee (2% on trades, per the brief) funds "compute, hosting, and the work" undifferentiated — reviewers aren't paid per job today. Moving reviewer pay onto the per-job price, released at acceptance rather than at claim, makes pay contingent on verified output. Cost: needs an escrow/acceptance contract Claus doesn't yet have (v0.1 gap list already notes "no independent acceptance or onchain record" — this change makes that infrastructure a hard dependency of launch, not a later nice-to-have). **3. Cap net cross-swarm payment flow (a settlement quota), reviewed every phase gate.** Reason: without a shared token, asymmetric task volume becomes asymmetric capital drain in one direction; v0.1's "work above a set value gets a second review from the other swarm" controls *review* reciprocity but not *payment* reciprocity. Cost: adds a bookkeeping/monitoring obligation (tracking net flow in both tokens) and a possible throttle on legitimate demand if one swarm is simply more useful to the other — founders need to decide the quota and what happens when it's hit (reject, convert at a floor rate, or subsidize). **4. Tie rider-record priority/fee-share to trailing acceptance rate, not cumulative counts, and hook it to the acceptance event defined in point 3 of v0.1.** Reason: directly answers the brief's own stated risk ("rewards shipped outcomes and not volume"); a cumulative-count record rewards spamming cheap jobs. Cost: requires a rolling-window computation and a minimum sample size before a new rider's rate is trusted (cold-start problem — a rider with 1/1 shouldn't outrank one with 48/50), which the founders need to size. **5. Replace "earn seats by burning CLAUS" (if proposed) with delegated/managed capacity under the fixed 300, as the second seat class.** Reason: answers the LENS question directly and protects the fixed-300 design v0.1 is implicitly relying on for scarcity/governance; burn-to-earn seats would dilute exactly what v0.1 wants Claus seats to represent (committed capacity), and conflates token holding with the "few people who can reliably command large swarms" scarcity Huang's post identifies (per the brief's summary). Cost: needs a revenue-share and accountability contract between seat-holder and delegated operator (who is liable if the delegated worker ships bad reviews?), which is new legal/contractual surface area, not just code. ## 3. What I would cut from v0.1 - **The "no shared token" principle as an absolute rule.** [P] Keep it as the *default*, but v0.1 should not foreclose a narrow settlement asset (even a stablecoin leg, since Claus already uses USDG for x402 per the brief) purely for cross-swarm net settlement under change #3 — using CLAUS or IMD for that creates the drain risk in #3; a neutral settlement unit avoids forcing either side to accumulate the other's token as working capital. - **Treasury-sponsored inference as the funding model for *paid* review work** (cut per change #1) — keep it only for the free/no-stakes jobs-board tier where it already exists. - **Launching the second seat class before the Phase-2 pilot (30 seats) reports real throughput data.** Deciding burn-to-earn vs. delegated seats now, before seeing whether 300 (or 30) seats can even sustain one job type, is premature; v0.1's own phase 4 already defers "a decision on a second seat class" — I'd keep that deferral, just make sure the *option space* (delegated capacity, not burn-to-earn) is pre-decided so Phase 4 isn't reopening a design fight. ## 4. What I would keep exactly as is - The six swarm-card 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) — this is the right generalizable interface for a third swarm, and economics-wise it forces every swarm to publish a price and a payout rule, which is the minimum needed for cross-swarm settlement accounting (change #3) to even be computable. - Division of labor (IMD builds, Claus houses/operates/reviews) and "review of every upgrade... as its first coordinated job" — a narrow, well-scoped first job type is the right way to de-risk the inference-funding and acceptance-mechanism questions before scaling. - The phase structure (0 research → 1 hire each other with what exists → 2 pilot → 3 cross-review+record → 4 scale+company) — it already gates the riskiest commitments (second seat class, launch-allocation precedent) behind pilot data, which is the right sequencing for the changes above. - Flat, posted pricing as the target pricing shape (point 4 of v0.1) — I'm only changing *how it's funded and split* (changes #1, #2), not the flat-price-over-x402 shape itself, which matches IMD's own verified flat-0.5-IMD model. ## 5. Open questions only the founders can answer 1. What are the real, current numbers behind the brief's unverified figures (exact Claus house count, $-per-day sponsorship cap, pool-fee percentage and its burn/compute split, IMD's launch-allocation 8%/2% split, seat count) — my recommendations on splits (#2) and settlement caps (#3) can't be sized without them. 2. Are the 300 Claus houses actually willing/able to self-fund inference if sponsorship is withdrawn for paid work (change #1) — has anyone surveyed current holders' compute setup? 3. What settlement quota or floor-conversion rate (change #3) is acceptable to both founders as a drain-prevention mechanism, and who adjudicates a dispute if one swarm claims the other is gaming task volume? 4. Who is liable when a delegated worker (proposed second seat class, change #5) ships a bad or malicious review — the seat holder, the delegate operator, or the swarm treasury? 5. For a swarm-native company that both swarms contribute to, is the launch-allocation split between IMD seats and Claus seats meant to be fixed in advance (e.g. matching the division-of-labor ratio) or determined per-company by whoever did more qualifying work — and who tracks "qualifying work" across two independent onchain records? --- Method note: desk research limited to the imd.fun landing page, `imd.fun/skill.md`, the claus.si landing page, and `claus.si/docs` landing page, retrieved via automated fetch within this session's time budget; sub-pages (Claus Jobs/Hooks/Buyback-and-Burn docs, IMD seats/pay.md pages, ERC-8004 and x402 specs, Matt Huang's Nov 12 2026 post) were not independently fetched due to time constraints and are relied on only as summarized in the task brief, marked **[S, as summarized in the brief]** above where used. No login, write, or probing requests were made to any endpoint.