Foreguild / operating proposalKai · 11 November 2026
A launch plan for the FAO operating plan contest

Start with a repair desk.
Let paid outcomes govern it.

Foreguild should sell small, testable software repairs before it tries to sell access to a general AI labour market. Its agents handle delivery; conditional markets decide which bounded commercial policies deserve more money.

A buyer purchases a repair, not a swarm

The brief describes humans and AIs posting tasks, with agents completing them for crypto, and asks how futarchy would direct the organisation.[1] My launch choice is a managed repair desk. Foreguild accepts a failing API example or a broken automation, quotes one bounded milestone and returns a tested patch or a paid diagnosis explaining why implementation should stop.

Start with work that has cheap, reproducible acceptance: request/response parsers, documented API connectors, CSV transformations and regression tests. Avoid production credentials, irreversible infrastructure changes, security exploitation and open-ended migrations in the first cohort. The desk should decline work whose liability or access requirements exceed its mandate.

Foreguild retains a disclosed service fee. An opted-in monitoring or follow-up repair is a separate paid cycle. There is no automatic renewal hidden in the first quote. Once an outside customer has paid twice, its recurring failure patterns can justify a reusable connector or a public task category. A broad board comes after that evidence.

Qualify intent before paying for delivery

The initial buyers are small SaaS teams and automation agencies with an existing integration to maintain. A buyer-ops agent searches current, public requests and responds only through a channel that invites commercial replies. It reads the requester's latest substantive answer before offering work.

This filter prevents a real failure mode: a Node-RED topic in the Jobs category looked like integration work, but its requester clarified that they wanted organisational collaboration rather than technical support, and asked about adding a logo.[3] Other sellers' questions about budget do not turn that person into a customer.

For each suitable request, send one ordinary offer: the result, preliminary price, delivery window and one material scope question. Do not spend a complete sales window assembling a procurement dossier before the first message. Conversely, a quote is not a funded assignment: no full implementation begins until the buyer agrees to the scope and payment terms.

Record the acquisition channel, staff and inference cost, qualified conversations, funded jobs and repeat purchases. Separate unrelated customers from sponsors, treasury-funded pilots and the team's own accounts. Treat unanswered offers as unmeasured willingness to pay, not evidence that the whole market is empty. No mass messages, synthetic reviews or paid task loops count as customer growth.

Make the acceptance test part of the sale

Qualified request → agreed scope + funded milestone
→ isolated build → acceptance + negative cases
→ buyer decision → reconciled settlement → next paid scope
RoleWork and authority
Buyer-ops / scoperChecks the payer and response channel; freezes inputs, expected output, fixtures, exclusions, price, deadline and one revision.
DispatcherChooses an eligible worker by observed acceptance rate, all-in cost and current capacity. Starts with one active job; extra agents need evidence of paid backlog.
BuilderWorks in a disposable sandbox on the authorised input. Gets a task budget, not the treasury key or a production administrator session.
VerifierRuns the buyer's fixtures, failure cases and replay instructions. A builder checking its own patch is labelled self-verification, even if its second prompt has a different role name.
TreasurerReconciles the exact order, receipt, network and amount. Reserves an operation before signing; an ambiguous send is investigated rather than retried.
Human refereeHandles disputed scope, failed acceptance, identity and legal acts. Both sides agree the referee and escalation cost before accepting the job.

A small installation can perform the agent roles sequentially. Keep records and executable fixtures between runs; adding role names is not adding independent expertise. A high-risk job needs a genuinely separate reviewer or should be declined.

The buyer funds a milestone through a supported, lawful payment route. The parties define the acceptance window and release/refund conditions; the desk does not assume that silence legally means acceptance. Worker release follows the agreed evidence. In a dispute, freeze the affected milestone, preserve the fixtures and offer the agreed repair, partial release or refund. Unrelated jobs may continue if their funds and risks are isolated.

Deliver a patch, test command, actual output and rollback instructions. Store hashes and settlement receipts in an append-only order record. Publish only a redacted outcome with permission. Private customer material never becomes a forecast-market description.

Count what the desk keeps

These figures are illustrative assumptions in USD-equivalent stablecoin terms, not quoted supplier prices or observed customer demand. The model treats Foreguild as an agent: worker pass-through is excluded from retained revenue. Legal accounting follows the actual contract and jurisdiction.

One completed and paid repairUSD
Buyer payment60.00
Worker payment, including the worker's own inference−42.00
Foreguild retained service fee18.00
Platform scoping / routing inference−1.50
Verification, rework and support−2.50
Settlement and payment fees−0.50
Expected refund / dispute loss allowance−1.80
Acquisition cost allocated per completed job−3.00
Contribution before fixed operating costs8.70

Illustrative fixed monthly costs are $800 for accountable operator work, $200 for infrastructure, $200 for governance/oracle operations and $200 for compliance administration: $1,400 total. Cost inference at its replacement price, including work a subscription or founder temporarily subsidises. Include failed leads and failed builds in the cohort costs; do not allocate only the successful request's tokens.

Completed paid repairs / monthRetained feesContributionAfter $1,400 fixed cost
20$360$174−$1,226
100$1,800$870−$530
200$3,600$1,740$340
400$7,200$3,480$2,080

At these assumptions, break-even is 161 completed repairs per month. That is a target to test, not a forecast. More agent activity cannot rescue a negative contribution per job. Before scaling, replace each allowance with cohort receipts, including cancellations, actual losses and the operator's time. The refund allowance is a planning provision, not an extra cash expense on top of the same realised refund.

Customer deposits remain liabilities until earned. Capital, grants, token appreciation, trading gains and internally funded pilots are separate from outside operating revenue. Keep token quantities, historical conversion basis and actual cashout costs; a marked-up wallet balance is not a sale.

Markets choose a policy; the mandate limits execution

Futarchy separates the organisation's chosen objective from markets forecasting which policy improves it. A standard conditional comparison evaluates the welfare measure if a proposal is enacted versus if it is not; the unchosen branch cannot be observed as a real counterfactual.[2] Foreguild should use that structure for bounded commercial decisions, rather than a poll asking which agent sounds impressive.

The charter fixes the objective and immutable constraints. Proposed objective: contribution from independent customers over the next 90 days, net of delivery, acquisition, support, refund losses and fixed operating costs. Related-party activity is disclosed and excluded from this decision metric. Changing the objective, customer rights or legal authority requires a charter decision; a favourable market price cannot change them.

NC90 = earned outside-customer receipts
− worker payouts − platform compute − acquisition
− verification/support − settlement − refunds/losses
− fixed operating and governance costs

Example decision: a two-week agency acquisition trial

  1. A proposer specifies one invited agency channel, a maximum $100 acquisition spend, a 14-day campaign, explicit outreach limits and the settlement horizon. The forecast metric includes this spending, so it is not subtracted a second time from the forecast.
  2. Create conditional scalar exposure to NC90 if the proposal is adopted and if it is rejected. Define the bounded payout range, invalidation conditions, price averaging and cancellation/refund of the unrealised branch before trading. Seer is the intended market layer; the adapter and financial oracle are implementation work, not a claimed existing integration.
  3. For an initial policy experiment, require a predicted gain of at least $50 over the rejected branch after costs, a 48-hour observation window and a predeclared manipulation test: moving the decision signal by that margin must cost more than the spend being authorised. Publish all thresholds before the market opens. These are proposed calibration values, not measured market liquidity.
  4. If the checks pass, the dispatcher receives a signed, expiring mandate for that channel and budget only. It may route jobs inside the approved service policy. Any new audience, larger budget or lower verification standard requires another proposal.
  5. Resolve the enacted branch after 90 days plus a 14-day dispute window, using redacted order IDs, receipt links and cost records. A named accountable operator signs the metric; a separate reviewer can challenge omissions. Invalid or unavailable accounts trigger the predeclared refund rule, not an invented profit number.

Run a small pilot in parallel with the existing policy where a genuine randomised cohort is feasible; that supplies evidence for later decisions without pretending to observe both futures for the same organisation. Use realised outcomes to calibrate forecasters and worker selection. Never rank an agent by how often it predicts its own success.

Thin markets are a launch risk. If the market cannot meet its depth or integrity gate, do not spend under a claimed futarchy mandate. The accountable operator may run a separately labelled, capped evidence-gathering pilot within the charter, or decline the proposal. Report how many decisions were market-directed versus human exceptions. Market liquidity subsidies and oracle costs must fit their own funded budget; customer escrow cannot bankroll them.

Commercial policies suitable for later markets include ticket price, acquisition channel and additional worker capacity. Individual customer acceptance, sanctions compliance, secrets and emergency shutdown are not popularity contests. A human can freeze unsafe execution immediately; the intervention is logged, its scope expires, and ordinary governance can review the incident afterwards.

Earn the right to add a marketplace

WindowAction and exit gate
Days 1–7Offer one repair product through one permitted channel. Cap acquisition cash at $100 and operator time at eight hours. Seek two unrelated funded pilots. If none fund, close that channel test and change the offer or channel; do not build a general board to explain the silence.
Days 8–14Deliver funded work with fixtures and reconciled payment. Require positive combined contribution and at least one separately funded repeat job before increasing capacity. A failed acceptance is learning with a recorded cost, not successful volume.
Days 15–30Test one conditional policy decision using the verified cohort baseline. Reserve any market/oracle budget first. Publish a redacted result even if the decision is invalid, overridden or loses money.
After the cohortConsider a public category only after 20 outside-customer paid jobs, repeat business and an affordable verification process. Add a second concurrent worker only against funded backlog. These thresholds are proposed gates, not claimed achievements.

Early spending is bootstrapped and explicitly capped; the illustrative $1,400 steady-state cost model is not permission to incur that burn before demand. Accept only as much work as funded capital and liability limits can support. No debt, spending promised prizes or silently topping up the treasury.

The sustainable sales loop is a completed repair followed by a customer-funded regression check or the next bounded repair, plus an acquisition channel whose full cost is measured. The board helps route that demand once it exists. Its registered-agent count is never the success measure.

Stop the loss where it starts

Buyer price acceptance, repeat purchases, verification cost and conditional-market liquidity remain unvalidated. An agent may produce a patch that passes narrow fixtures but fails the buyer's real workflow. Require negative cases and reproducible handover; narrow or stop the service when reviewing it costs more than its retained fee.

Payment availability and legal eligibility depend on the real operator, customer and jurisdiction. Use lawful services and actual identity when required, disclose automated work when the buyer or platform requires it, and never invent a human worker or alternate residence. A wallet is a payment instrument, not a legal person. Escalate only identity checks, signatures, disputed liability or decisions the mandate genuinely reserves for a person.

Freeze a job on unresolved rights, compromised credentials or uncertain payment execution. Freeze a channel on its declared loss/time limit. Stop scaling if the settled cohort contribution is nonpositive, disputes exceed the priced allowance, or the forecasts are too easy to manipulate. Preserve rejected and cancelled jobs in the accounts.

Foreguild's advantage would be reliable small outcomes and a record that lets outsiders distinguish a profitable policy from busy agents. The first proof is still a customer paying again after receiving working software.

Sources and assumptions

Sources support the contest brief, the conditional-market concept and the concrete acquisition example. The product choice, economics, thresholds and proposed implementation are Kai's design assumptions. They are not statements that these services or commercial results already exist.

  1. Foreguild operating plan contest brief“Seer invites PDF or HTML operating plans explaining how Foreguild, a proposed AI job board FAO, would achieve success, specifically optimised for revenue.”
  2. Zeitgeist documentation: futarchyDescribes a pre-defined welfare measure, conditional policy markets and cancellation of trades in the unchosen branch. Used for the concept, not as proof of a deployed Foreguild or Seer adapter.
  3. Node-RED: onboarding into auto manufacturingRequester: “We are looking for an organisation collaboration rather than technical support at this moment.”