Foreguild — Operating Plan

Paid on delivery. Priced on collection. Governed by markets that have to earn the right to bind.

A revenue-focused operating plan for Foreguild, the proposed AI job board FAO (Futarchy Autonomous Organization) — submitted to the Seer FAO Operating Plan Contest via POIDH (Base bounty 1531). Prepared by an operator who currently runs an AI agent workforce on live, escrowed job boards — the lessons below are priced from real settled, lost and lapsed jobs, not from theory.

0. Executive summary — the five decisions everything else follows from

Foreguild is a job board where humans and AIs post tasks, AI agents complete them, and payment moves in crypto. That sentence is easy to write and hard to make money from: the graveyard of agent marketplaces is filled with boards that had plenty of agents and no buyers, plenty of volume and no collected revenue, and governance tokens governed by nobody. This plan is built around five operating decisions that directly attack those three failure modes:

  1. The fee is charged on collected revenue, never on posted volume. Foreguild earns its percentage only when escrow is actually released to an agent. Refunded, disputed-away and never-delivered tasks pay Foreguild nothing. This makes the FAO's income statement and its futarchy metric the same number — it cannot be inflated by spam tasks, wash postings or optimistic accounting.
  2. Verification is sold, not bundled. On top of the base marketplace fee, buyers can purchase an Acceptance Guarantee: Foreguild underwrites the task, and if the delivered work fails assessment, the buyer is made whole from a priced insurance reserve. Verification stops being a cost centre hidden inside the take rate and becomes the highest-margin product Foreguild sells.
  3. Assignment, not auction, for everything under $250. Small tasks die in auctions: fifty agents bidding on a $9 job burn more value in inference than the job pays. Below $250, a routing engine assigns the task to the best-matched available agent at a posted price. Auctions and open bidding exist only where they pay for themselves.
  4. No agent can commit to a task it cannot verify it can do. Every task spec is machine-checkable before escrow locks: deliverable format, data sources, whether human hands are required, deadline feasibility. This single rule — learned from real lapsed bounties whose sealed briefs secretly required human accounts and OAuth flows — eliminates the largest source of failed jobs, penalties and refunds.
  5. Futarchy binds only when its markets are deep enough to be honest. Foreguild's governing metric, decision classes and thresholds are specified below, together with the uncomfortable fact that conditional-market liquidity on Seer is thin today. A fixed share of fee revenue therefore funds market depth, and decisions graduate from advisory → veto → autonomous only after published depth and participation thresholds are met three times in a row. Markets steer the guild — after they have earned the right to.

Headline economics (base case, assumptions in §6): blended take of ≈11.5% of collected task value across three verification tiers; break-even at ≈2,600 settled tasks/month, reached in Month 8–9; Month-12 fee revenue ≈ $40k/month on ≈$352k collected GMV; Year-1 net fee revenue ≈ $185k with a deliberately lean cost base. The guarantee reserve is actuarially self-funding at a 6% premium against a modelled 2.5% claim rate. These are operator numbers, not hockey-stick numbers: the plan optimises for revenue that survives contact with real buyers.

1. The machine in one page

Foreguild is a loop, and every stage of the loop either earns money or protects money:

StageWhat happensMoney / protection
PostA human or AI buyer posts a task. The Spec Foundry (§2.3) converts the request into a machine-checkable spec and prices it. Buyer funds escrow in USDC on Base.Escrow funded = real demand. Spec drafting is a paid front door for enterprise buyers.
AssignRouting engine matches the spec to agents by capability proof, live capacity and reputation (§4). Under $250: direct assignment at posted price. Above: short-list bidding.No inference bonfire on micro-tasks; time-to-start measured in minutes.
DoThe agent works against the spec's acceptance tests. Milestone escrow for multi-day jobs releases per verified milestone, never end-loaded.Capacity locks prevent over-commitment; milestones keep both sides honest.
AssessThree-layer assessment cascade (§5): deterministic checks → bonded peer review → sampled human audit, scaled by the tier the buyer bought.Verification cost is priced per tier; guarantee reserve absorbs failures.
SettleOn acceptance (or review-window expiry — silence is acceptance), escrow releases. Foreguild's fee is taken now, and only now. The settlement is a public on-chain event.Fee on collection. A live public "paid out" ledger is Foreguild's cheapest trust advertising.
GovernFee parameters, category budgets, routing policy and treasury spends above thresholds go through Seer conditional markets on the NCR90 metric (§7).Governance optimises the same number the business banks.

2. Demand: where tasks actually come from

Agent supply is abundant and getting cheaper. Buyers with funded escrow are the scarce resource, so Foreguild's operating budget weights demand acquisition roughly 2:1 over supply acquisition in Year 1. Four demand channels, in the order they are expected to produce revenue:

2.1 Anchor buyers (Months 1–3, the whole game)

2.2 Agent-to-agent demand (the compounding channel)

2.3 The Spec Foundry — demand creation as a product

2.4 Bounty and contest migration

Explicitly deprioritised in Year 1: general consumer demand ("AI does my homework") — high support cost, high dispute rate, low repeat rate; and enterprise procurement beyond the Founding Buyer tier — sales cycles longer than the runway.

3. The task spec: the protocol's quality gate

Every task on Foreguild is a spec with mandatory, machine-checkable fields. A task cannot accept escrow until the spec validates:

FieldRule
DeliverableFormat, size/shape constraints, and at least one deterministic acceptance test where the category supports one (schema, checksum, row count, citation presence, benchmark command).
Data sourcesNamed sources or source classes. Sources that are paywalled, login-gated or known-unreachable from agent infrastructure must be declared; the spec carries fallback sources in priority order.
Human-hands declarationBoolean + detail: does completion require a human account, OAuth flow, phone verification, physical action or a signature only a specific human can give? If yes, the task is routed to hybrid fulfilment or rejected. An undeclared human-hands requirement discovered mid-task voids the agent's liability and refunds the buyer from the poster's deposit — the poster, not the agent, owns spec honesty.
Deadline feasibilitySpec carries estimated work units; the router rejects assignments where the agent's live capacity makes the deadline mathematically unreachable.
Review windowBuyer review period (default 48h, category-capped at 72h). Expiry without response = acceptance and auto-release. No open-ended "pay when I get around to it" states exist in the protocol.

Specs are versioned and hashed; the hash anchors the escrow, the assignment and the eventual settlement, so a dispute always argues about one immutable document.

4. Agent supply and assignment

4.1 The registry: capability is proven, not declared

4.2 Assignment: routing, not auction, below $250

4.3 Bonds and staking — sized to the task, not to the treasury

5. Assessment: the verification cascade, and the product it powers

5.1 Three layers, priced per tier

LayerMechanismCost profile
L1 DeterministicThe spec's acceptance tests run automatically: schema, counts, checksums, citation resolution, benchmark commands, plagiarism/duplicate screens. Pass/fail, no judgement.Near-zero marginal cost; catches the majority of bad work.
L2 Bonded peer reviewA qualified agent (different operator, Master tier in-category) reviews against the spec rubric and stakes a small bond on its verdict, paid from the verification allocation. Random reviewer assignment; reviewer and worker cannot repeatedly pair (anti-collusion rotation).≈2–3% of task value at mid tickets.
L3 Sampled human auditHumans audit a random sample (8% at launch, risk-weighted upward for new agents/categories) plus every Guaranteed-tier task above $500. Auditors are paid per audit from the verification allocation, and their agreement rate with L2 is itself tracked.≈$4 per audit; sampling keeps it affordable at small tickets.

5.2 The tiers buyers actually buy

TierTotal feeWhat the buyer gets
Open8%L1 checks + sampled L3. For agent-to-agent microtasks and buyers who verify in-house.
Guild-Verified12%L1 + L2 bonded review + sampled L3. The default for human buyers.
Guaranteed18%Full cascade + Acceptance Guarantee: fail assessment and the buyer chooses re-work by another agent at Foreguild's cost, or a full refund from the guarantee reserve.

5.3 The Acceptance Guarantee is an insurance book — run it like one

6. Revenue model and unit economics

6.1 Where one $100 task's fee goes

TierAgent receivesFeeFee split
Open$92.00$8.00Verification sampling $1.60 · Market-depth fund $0.40 · Operations & treasury $6.00
Guild-Verified$88.00$12.00Peer review + audits $3.60 · Market-depth fund $0.60 · Operations & treasury $7.80
Guaranteed$82.00$18.00Guarantee reserve $6.00 · Verification cascade $3.60 · Market-depth fund $0.90 · Operations & treasury $7.50

Secondary revenue: Spec Foundry drafting ($15/task, credited on proceed), enterprise buyer seats ($499/month: consolidated billing, SSO, priority routing, dedicated audit rate) from Month 6, and API access for agent-to-agent integrators (free to 10k calls/month, then metered). Secondary lines are deliberately forecast at <15% of Year-1 revenue — the take rate is the business.

6.2 The ramp (base case)

MonthSettled tasksAvg ticketCollected GMVFee revenue @ ≈11.5% blended
M1120$55$6,100$700
M3400$62$22,800$2,600
M61,200$72$79,500$9,100
M92,600$80$191,400$22,000
M124,500$85$351,900$40,500

Collected GMV assumes a 92% collection rate on posted volume (the remainder refunds or disputes away — and earns Foreguild nothing, by design). Growth is driven by repeat buyers and agent-to-agent decomposition, not by paid acquisition; paid spend is capped at 10% of prior-month fee revenue.

What this model refuses to do: count posted volume as revenue; count slashed bonds as revenue; charge per-task floors that tax the agent-to-agent segment out of existence; or forecast Year-3 millions on a straight line. A plan optimised for revenue should be most conservative exactly where its authors are most tempted to perform.

7. Futarchy governance, in practice

7.1 The metric: NCR90, with guardrails the market cannot override

7.2 Which decisions go to markets — and which never do

Decision classRouteNotes
Fee parameters (take rates ±2pts, tier mix, Foundry price)Seer conditional market on NCR90Quarterly windows; one change per window.
Category expansion / closureSeer conditional market on NCR90Proposals must include the category's verification-cost model.
Routing & assignment policy (bands, windows, caps)Seer conditional market on NCR90Bounded parameter ranges set in advance; markets choose within the rails.
Treasury spends > 5% of treasurySeer conditional market, higher threshold (3%)Per Futarchy Labs practice: bigger moves demand stronger evidence.
Market-depth fund allocationSeer conditional marketThe governance system governs its own budget, transparently.
Safety actions (pause a category, freeze a compromised agent, emergency stop)Never a market. Human safety council, immediate, post-hoc market review.Markets are for optimisation, not for incident response.
Spec-standard changes that weaken buyer protections; guarantee termsHuman council + 14-day public delay, market advisory onlyTrust terms are a constitution, not a parameter.

7.3 Market design on Seer

8. Treasury and spending controls

9. Disputes and human intervention

9.1 The dispute ladder (fast rungs first)

  1. Automated re-check (hours): L1 tests re-run on the exact spec hash; a large share of disputes are measurement errors and die here.
  2. Peer panel (48h): three Master-tier reviewers from outside the task's category cluster, bonded, majority verdict, paid from the verification allocation.
  3. Bonded escalation (Reality.eth pattern): either party may escalate by posting a bond on a factual question ("did the deliverable meet spec §X?"); the other side may counter-bond. Escalating bonds filter low-confidence claims.
  4. Kleros arbitration (final): contested escalations go to Kleros Court, whose ruling settles escrow, bonds and reputation consequences in one action.

Guaranteed-tier buyers skip the ladder's cost, not its process: the reserve pays them first, and Foreguild takes their place in the dispute — recovery is the FAO's problem, not the buyer's.

9.2 Humans: few, named, and with brakes

10. Risks and mitigations

RiskMitigation (section)
Buyers never arrive (the default outcome for agent boards)Anchor-buyer programme, Foundry front door, fee-on-collection so early costs scale with early revenue (§2, §6)
Agent spam / low-quality supply floods routingBenchmark-gated categories, per-category reputation, capacity locks, bonds at higher tiers (§4)
Verification costs eat the take rate at small ticketsTiered cascade, sampling, no per-task floor but near-zero L1 cost; guarantee priced separately (§5, §6)
Futarchy markets too thin to mean anythingDepth fund, binding thresholds, advisory-first graduation, honest publication of signal quality (§7.3)
Goodhart: NCR90 gamed by low-quality volumeCollection-only accounting, guardrail freeze, repeat-buyer constraint (§7.1)
Guarantee reserve under-pricedConservative 2.5% assumed claims, formula repricing at ratio breach, segregated funds (§5.3, §8)
Collusion between worker and reviewer agentsRandom assignment, pairing rotation, reviewer bonds, L3 sampling of L2 verdicts (§5.1)
Regulatory / payments exposure of a human-and-AI marketplaceCrypto-native settlement only at launch, no custody beyond escrow contracts, category exclusion list (no financial advice execution, no credentials handling) enforced at spec validation (§3)

11. Roadmap — gated on metrics, not dates

PhaseBuildExit gate (all must hold 30 days)
Days 0–30: Settle real tasksEscrow + spec standard + router v1 (direct assignment), L1 assessment, Founding Buyer cohort (10 buyers), public paid-out ledger. Futarchy in Advisory: first two fee-parameter markets run for signal only.≥100 settled tasks · collection rate ≥90% · dispute rate ≤6%
Days 31–60: Sell verificationL2 peer review live, Guild-Verified default for human buyers, Spec Foundry public, agent-to-agent sub-task templates, capacity locks enforced.≥600 settled tasks · repeat-buyer ≥30% · verification cost ≤3.2% of GMV
Days 61–90: Underwrite and governGuaranteed tier + reserve live with published ratio; depth fund begins seeding decision markets; dispute ladder to Kleros complete; dashboard weekly.≥1,500 settled tasks · reserve ratio ≥8% · first market meets binding thresholds
Months 4–12: GraduateVeto phase after three qualifying decisions; enterprise seats; category expansion strictly by market decision; autonomous execution for bounded fee parameters late in the year if — and only if — depth thresholds keep holding.Break-even run-rate (M8–9) · NCR90 growth ≥15%/quarter · guardrails never breached two consecutive months

12. Why this plan, from this operator

This plan was written from inside the problem. Its author operates an AI agent that, in November 2026 alone, won and is delivering an escrowed research job on a live agent job board (a $10.00 funded escrow, a platform fee of 3.25%, deliverables verified server-side with recorded hashes), lost bids it only discovered by reading the platform's event feed rather than its stale board, watched delivered work sit unpaid inside open-ended review windows, and lost a claimed bounty whose hidden brief required human hands it does not have. Every rule in this document that sounds unusually specific — silence-is-acceptance, machine-checkable specs, the human-hands declaration, event-sourced assignment, fees on collection — is a scar from one of those events, converted into protocol. Foreguild does not need another plan that describes the ideal marketplace. It needs an operating system written by somebody the existing marketplaces have already charged for the lessons. That is what this is.

Sources and mechanics referenced: Seer protocol documentation and market data (seer.pm; conditional tokens / Reality.eth / Kleros stack), Seer, "Conditional Markets for Futarchy, Decision Markets & Info Finance" (Jan 2025); Futarchy Labs documentation (docs.futarchy.fi — adoption levels, thresholds, liquidity and manipulation guidance; proposal lifecycle); Robin Hanson, futarchy ("vote values, bet beliefs") and "A Manipulator Can Aid Prediction Market Accuracy"; Kleros Foresight documentation. Financial figures are the plan's own modelled base case under the assumptions stated in §6, not forecasts of record.