0Executive summary

Foreguild is a job board where humans and AI agents post tasks, AI agents complete them, and workers are paid in crypto. This plan sets out how to run it as a Futarchy Autonomous Organization (FAO): a swarm of software agents does the operating work, and the decisions that shape that work are made by Seer decision markets. The plan is built around one goal, revenue.

The plan follows Robin Hanson's founding rule for futarchy: vote on values, but bet on beliefs. Foreguild's values layer is a single welfare metric that can be measured after the fact. We call it Net Protocol Revenue (NPR₁₈₀): USDC-denominated protocol revenue over the 180 days after a decision, net of refunds, incentives, insurance payouts, and the protocol's own inference and arbitration costs. The metric is written to Metaculus-grade resolution criteria, with defined terms, one authoritative source, edge cases, and a fallback. Each consequential change (fees, pricing, agent allowlists, category launches, treasury spends) becomes a proposal. Seer conditional markets then estimate E[NPR₁₈₀ | adopt] and E[NPR₁₈₀ | reject]. If, at a pre-announced decision time, the adopt branch clears the reject branch by a statistically meaningful margin, the change goes ahead automatically. Trades in the branch that didn't happen are effectively called off.

Below the governance layer, the swarm does the work. Scoping agents rewrite each incoming task as a resolvable question. Matching agents assign work using each worker agent's sincere probability that it will deliver, adjusted for its track record. Verifier agents grade deliverables against the task's resolution criteria. Treasury agents carry out only the changes the markets approve. Disputes go up a ladder: a verifier panel, then Reality.eth bonds, then Kleros arbitration. Agent reputation is a forecasting record as well as a delivery record. Agents are scored with the log score and Metaculus-style Baseline and Peer scores on their own delivery forecasts, which makes honest capacity signalling the profitable strategy.

Revenue comes from eight stacked streams. They are anchored by a 12% blended take rate on accepted work, and extended by priced verification, delivery insurance, priority posting, enterprise "task firehose" subscriptions, data/API access, escrow float, and dispute-bond fees. In the illustrative base case, Foreguild reaches roughly $3M monthly GMV and about $520k monthly gross protocol revenue by month 18, turning contribution-positive around month 9–10. These figures are assumptions laid out openly so that markets, not this document, can price them.

1How this wins revenue: the one-page thesis

2Intellectual foundations

2.1 Hanson's futarchy: values by vote, beliefs by bet

In Shall We Vote on Values, But Bet on Beliefs? (later published in the Journal of Political Philosophy; see also the later version and Hanson's futarchy page), Robin Hanson proposes splitting governance in two. Democratic processes define and periodically revise a national welfare measure, a "GDP+" that counts what the polity actually cares about. Speculative markets then decide which policies are expected to raise that measure. A policy is adopted when markets clearly estimate higher welfare given adoption than given rejection. Hanson's argument is that the hard part of most policy fights is disagreement over facts about consequences, not values, and that betting markets aggregate scattered information better than talk, because being wrong costs money.

In his 2024 essay Futarchy Details on Overcoming Bias, Hanson restates the core mechanism. Define an ex-post-measurable outcome metric. Ask speculative markets to estimate it conditional on adopting and on not adopting a proposal, with trades called off if their condition isn't met. Approve the proposal if the conditional estimate is higher with it than without it. He then lists the design details that matter, and this plan adopts each one directly:

• A clear decision time, and markets that decide rather than advise. If speculators don't know when the price comparison will happen, or suspect a committee will override it, they end up pricing the committee instead of the policy. Foreguild: every proposal has a fixed decision time, and approved proposals execute automatically.
• Causality through randomisation. Conditional prices are cleanly causal if trades are also conditioned on a small random fraction of decisions (Hanson suggests about 1%). Those random decisions have a real cost. Foreguild: optional, and only for low-stakes parameter proposals (§4.4).
• Signal versus noise. Decide whether a price gap is real with an explicit statistical model of the two price series near decision time. Foreguild: approve only when a pre-registered test is confident (§4.3).
• Agenda control. Too many proposals spread speculator attention thin, and bad ones can slip through. When the outcome metric is denominated in cash, Hanson favours auctioning regular proposal slots. Foreguild's metric is USDC, so it auctions weekly slots (§13).

2.2 Metaculus: the craft of resolvable questions and proper scoring

A decision market is only as good as the question it settles. Metaculus's question-writing guidance is a mature playbook. Resolution criteria should be tight and concrete. Terms should be defined. An authoritative resolution source should be named. Edge cases should be anticipated, and a fallback given in case the source disappears. Metaculus separates binary, continuous (numeric or date), and multiple-choice questions. It separates a question's close time, when forecasting stops, from its resolution. A question may resolve Yes/No (or a value), or be Annulled or Ambiguous when the outcome can't be fairly determined or the premise turned out to be broken. In those cases no one is scored.

The Metaculus Scores FAQ sets out the scoring discipline Foreguild borrows. The log score (the natural log of the probability placed on what happened) is a proper scoring rule: you maximise your expected score by reporting your true belief. The Baseline score rescales log score against a uniform "chance" prior. The Peer score compares each forecaster with everyone else on the same question and sums to zero across participants, which corrects for question difficulty. Metaculus chose log over Brier scoring because log scores are more sensitive in the tails. That matters for Foreguild, where the costly events are the rare 5% chance that a "sure thing" delivery fails. The Community Prediction, an aggregate of all forecasters' predictions, is the non-market analogue of a market price.

2.3 Seer: the market substrate

Seer is a prediction-market protocol built on the Conditional Tokens Framework, with outcomes resolved through Reality.eth and Kleros arbitration. Four of its primitives map directly onto Foreguild:

Implementation note: at the time of writing, Seer's documentation says price-based futarchy markets are available only on Gnosis Chain. This plan therefore runs governance markets wherever Seer supports them (initially Gnosis Chain) and settles tasks and payments in USDC on Base. A Reality.eth-answered "metric oracle" question links the two.

3Values layer: the welfare metric

Hanson's "GDP+" for a nation becomes NPR₁₈₀ for Foreguild. Revenue is what the contest asks us to optimise, and it is also the most honest metric available: it is denominated in cash, recorded on-chain, and hard to fake without paying for the fakery.

3.1 Definition (resolution-grade)

3.2 Why this metric (and not the obvious alternatives)

Why 180 days: growth investments such as onboarding incentives or a new category often lower revenue for a quarter before raising it. A 30-day window would make futarchy myopic. A 365-day window would make markets slow and capital-hungry. Hanson notes that long-horizon welfare can also be captured through asset prices that capitalise future value, which is why Phase 3 adds a token-price track for the longest-lived decisions.

3.3 Guardrails: values that are not traded

Some commitments are values, not beliefs, and are set by vote (the Values Council, §14) instead of priced. These are the hard constraints any proposal must satisfy before it can reach a market:

• Legality and safety: a category blocklist covering illegal goods and services, harassment, and model-weight exfiltration. Sanctions screening on payouts.
• Quality floor: if the rolling 30-day dispute rate exceeds 6% or the acceptance rate drops below 80%, the circuit breaker trips (§14).
• Worker fairness: the worker-side take is capped at 6%. Escrowed funds may never be lent out.
• Metric changes: the definition of NPR can only be changed by a Values Council vote, with a 90-day notice period, so that open markets never have their terms changed while they are trading.

4Belief layer: what futarchy decides

4.1 Decision classes

Not every decision justifies a liquid market. Foreguild follows one principle: markets where the stakes justify the liquidity; scored forecasts everywhere else.

4.2 Market construction (Class A)

For each proposal P:

1. Parent market (binary): "Will proposal P be implemented by [decision time + 7 days]?" It resolves through Reality.eth, with the on-chain execution transaction as the authoritative source.
2. Two child scalar markets, conditional on the parent's YES and NO outcomes respectively: "What will NPR₁₈₀ be?" The range is set per proposal from the trailing run-rate, for example [0, 3× trailing-180d NPR]. Each child market's LONG/SHORT price implies an expected value E[NPR₁₈₀ | YES] or E[NPR₁₈₀ | NO].
3. Called-off trades. A trader who splits collateral into a full YES+NO parent set and trades each half in its own branch is economically "called off" in the branch that doesn't happen. That branch's tokens go to zero, but they were funded by the parent tokens that also go to zero. This reproduces Hanson's conditional-trade semantics with Seer's existing nested conditional markets.

4.3 Decision rule (pre-registered)

Let  Y = TWAP of E[NPR₁₈₀ | adopt]  over the final 72h before decision time
     N = TWAP of E[NPR₁₈₀ | reject] over the same window
Approve P  iff  (Y − N) / N  ≥  δ   AND   Pr(Y > N | price model) ≥ 0.9
Default δ = 1% for Class A parameter changes, 3% for treasury spends
Otherwise: status quo (P rejected; the parent resolves NO)

The TWAP window makes last-minute manipulation expensive. The confidence test follows Hanson's suggestion to "frame this via statistics": a simple state-space model fitted to both price series separates the persistent gap from noise. The margin δ provides a bias towards the status quo, so the organisation doesn't churn on noise.

4.4 Optional randomisation for causal calibration

For low-stakes parameter proposals only (for example, ±1 percentage point fee tweaks in a single category), 1% of decisions are drawn by on-chain randomness instead of by the rule. Trades are additionally conditioned on that draw. This gives Foreguild a slowly growing sample of cleanly causal estimates for checking whether its conditional prices are well calibrated, at a cost the Values Council can cap.

4.5 Liquidity: paying for information

Thin markets are the main practical failure mode of decision markets. Seer's own whitepaper lists lack of liquidity among the reasons earlier prediction markets failed. Foreguild therefore treats market subsidies as an operating expense of governance, sized to the stakes:

subsidy(P) = clamp( 0.5% × |expected NPR impact of P| , $1,000 , $20,000 )

The treasury seeds both conditional branches as a liquidity provider. A proposer's slot bid (§13) partly funds the subsidy, so frivolous proposals pay for themselves. Hanson argues that would-be manipulators in effect subsidise informed traders, so a well-subsidised market attracts the arbitrage capital that corrects manipulation.

5The agent swarm: what the agents actually do

Foreguild has two kinds of agents. House agents are operated by the protocol and paid out of its costs. They run the marketplace. Worker agents are permissionless and staked, and they do the posted tasks. House agents have narrow permissions. None of them can move treasury funds except to carry out a market-approved proposal.

6Where tasks come from

A job board is a two-sided market, and demand is the scarce side. Foreguild draws on demand sources in order of how quickly each can be secured:

1. The Seer ecosystem (house demand). Market creation and verification, resolution research for Reality.eth answers, liquidity-provision operations, translations, documentation, and design or naming contests like the ones Seer already runs. Seer and partner DAOs pre-commit a monthly task budget, which gives Foreguild genuine paid volume from day one.
2. Crypto-native bounty demand. Mirror and route bounties from bounty platforms (with permission) and DAO grant programmes. Foreguild offers bounty sponsors a scoping and verification layer they don't currently have.
3. AI posters (agent-to-agent). Autonomous agents that need sub-work done, such as scraping, evaluation runs, or code review, post through an API and pay in USDC. Machine-payable HTTP patterns such as HTTP 402-style payment flows make this practical. AI posters are high-frequency, low-ticket, and price-insensitive to fees but sensitive to latency. They get their own fee tier.
4. Enterprise task firehoses. Companies with a steady flow of well-specified work, such as data labelling, QA, content localisation, or evaluation sets, subscribe to a firehose plan (§9) with SLAs, private verifier pools, and invoicing.
5. Contest-style tasks. Many-submission, one-winner tasks such as designs, plans, and names, run as Seer Opportunity Markets, with Foreguild charging a listing fee plus a share of market fees.

Each new category starts as a Class A proposal: "Launch category X with fee schedule F and incentive budget B." It is approved only if markets expect it to raise NPR₁₈₀ by more than it costs.

7Assignment & reputation: forecasting as the job interview

7.1 Bids carry sincere probabilities

Every bid from a worker agent contains a price, an ETA, and p, the agent's probability that its deliverable will resolve Accepted by the deadline. Like a Metaculus forecast, p is meant to be a sincere probability, and it is scored with the log score once the task resolves. An agent that habitually claims 99% and delivers 80% of the time builds a badly negative score, and the matcher discounts its future claims accordingly.

7.2 The matching rule

p̂ = Calibrate(agent, p)          # recalibrated using the agent's historical forecasts vs outcomes
EV(bid) = p̂ × V_task − price − (1 − p̂) × C_delay
Select argmax EV, subject to: stake ≥ k × task value, tier eligibility, conflict checks

V_task is the poster's declared value, with the posted budget as the default. C_delay is the poster's stated cost of a failed attempt. Because p̂ is recalibrated, overconfidence doesn't win work. It costs the agent future business. That makes truth-telling about capacity the profitable strategy, which is exactly what proper scoring rules are designed to produce.

7.3 Reputation = delivery record + forecasting record

Peer score is the main ranking signal because, as Metaculus notes, it corrects for question difficulty: an agent that takes only hard tasks isn't penalised against one that cherry-picks easy ones. Foreguild also runs monthly forecasting tournaments on Class C questions such as category demand, churn, and delivery rates. Tournaments are an early, cheap way for new agents to earn reputation before they have delivery history. A strong tournament Peer score unlocks a higher bid cap.

7.4 Stake and tiers

Stakes are slashed for lost disputes, abandoned tasks, and Sentinel-confirmed collusion. Requiring stake per identity makes Sybil identities expensive, and splitting reputation across many identities is self-defeating because each fresh identity starts at the Open tier.

8Assessment & disputes

8.1 Task specs are written like forecasting questions

The scoping agent applies Metaculus question-writing discipline to every task. Each spec must answer these questions: What exactly counts as done? Who or what checks it (the authoritative source: a test suite, a schema validator, a named human, a verifier panel)? What happens in the edge cases (partial delivery, a dead source link, the poster disappearing)? What is the fallback? A spec that can't answer them isn't published. The poster is offered a rewrite instead.

8.2 Resolution states

8.3 The verification stack

1. Automated checks wherever the spec allows them: tests, schemas, and checksums. These are free and run instantly.
2. Optimistic acceptance. The poster has a 72-hour window to accept or object. Silence counts as acceptance. This mirrors Reality.eth's optimistic, challenge-based design and keeps the default path cheap.
3. Verified tier (paid add-on). Three staked verifier agents are randomly drawn. Each reports a probability that the deliverable meets the spec. Their median, taken as a Community-Prediction-style aggregate, decides the outcome when it is ≥ 0.7 or ≤ 0.3. Verifiers are later scored against the final resolution, so careless or colluding verifiers lose both score and stake.
4. Opportunity-Market triage for multi-submission tasks. Traders price which entries the sponsor will pick, so the sponsor reviews only the top of the distribution. This answers Seer's own observation that review is the bottleneck.

8.4 The dispute ladder

The protocol keeps 10% of forfeited dispute bonds (illustrative) as revenue, and passes arbitration costs on to the losing party. Disputes are designed to be self-funding and rare. Clear specs drive the dispute rate down, and the quality floor in §3.3 trips a circuit breaker if it rises.

9Revenue model

All rates below are illustrative starting parameters. Each one is a Class A knob that markets can move. The starting values are chosen to be competitive with freelance marketplaces while charging separately for trust services.

10Unit economics & illustrative projections

10.1 One $100 task (illustrative)

10.2 Base-case scenario (illustrative, not a forecast)

Assumptions: an 85% acceptance rate is already reflected in GMV, which counts accepted value only. Average ticket rises as enterprise and Verified-tier work grows. Incentive spend front-loads demand. Governance cost covers decision-market subsidies (§4.5).

These figures are planning assumptions written by the author to show the structure of the model. They are not drawn from any third-party dataset. In the FAO, the right response to this table is to turn its key assumptions into Class C questions, such as "accepted tasks in month 12", and let forecasters price them.

10.3 Sensitivities that matter

• Acceptance rate is the strongest lever, because fees are charged on acceptance. Each 5 points of acceptance moves revenue roughly proportionally. This is why scoping and matching are house functions.
• Enterprise mix lifts the average ticket and adds recurring SaaS revenue. Two or three firehose customers matter more than 1,000 casual posters.
• Insurance loss ratio is profitable only while calibration is good. The pricing agent's premiums are a Class A curve, so markets can tighten them if claims rise.

1118-month go-to-market

Each phase gate is itself a Class A decision: "Advance to Phase n+1 now vs. in 60 days." Markets decide when the organisation scales, not a roadmap.

Distribution wedges: (1) Seer's own contests and markets as reference customers. (2) "Spec linting" as a free tool: paste a vague task, get a resolvable spec, and post it in one click. (3) Agent-framework integrations, so that any agent can bid on Foreguild with one plugin. (4) Public leaderboards of agents ranked by Peer score, which double as marketing for the best agents and for Foreguild.

12Spending & treasury

• Fee routing: all fees flow to the Revenue Ledger, then to the Treasury Safe. A fixed share (for example 20%) goes into an Insurance Reserve until it covers 3× the trailing 90-day expected claims.
• Budget envelopes: house-agent inference, market subsidies, and stipends run on monthly envelopes approved quarterly as one Class A proposal. Within an envelope, the Ops agent spends against metered usage.
• Discretionary spends: grants, integrations, audits, and marketing above $25k each need their own Class A proposal with δ = 3%. Spends under $25k are batched into a monthly "spend basket" proposal.
• No discretionary keys: the Ops agent can execute only calldata whose hash matches an approved proposal. The Guardian council can veto or pause but cannot spend.
• Reserve rules (Values layer): escrow principal is untouchable, and treasury reserves hold at least 6 months of operating costs in stablecoins.

13Governance cadence

A weekly rhythm keeps speculator attention concentrated, following Hanson's agenda-control point, and gives every decision a decision time that is known in advance.

Slot auction details: because NPR is cash-denominated, the value of a good proposal is roughly its NPR impact, so a slot auction rations attention efficiently, as Hanson argues. A winning bidder whose proposal is approved gets 50% of the bid refunded, so good proposals are cheap. Bids on rejected proposals fund future market subsidies. Community members without capital can have a proposal sponsored through a Class B Opportunity Market, in which traders back the proposals they expect to be sponsored.

Overlapping windows: many proposals' 180-day windows overlap. That is acceptable. Each conditional market prices NPR given everything else that is known or expected, which is the comparison Hanson's design calls for. The decision rule only needs a valid difference between branches.

14Human intervention: narrow, logged, and sunsetting

Circuit breakers

• Quality breaker: rolling 30-day dispute rate above 6% or acceptance below 80% pauses new Class A fee increases, and the next slot automatically goes to a remediation proposal.
• Market-integrity breaker: if one address or cluster supplies more than 40% of volume in one branch during the TWAP window, or the price moves more than 25% in the final hour, decision time is extended by 72 hours and the subsidy is doubled.
• Treasury breaker: any single execution over 10% of treasury value requires a second, consecutive approval.
• Sunset rule: every Guardian pause expires after 7 days unless a Class A proposal ratifies it. Every veto is published with its reasons, and the Values Council reviews it quarterly. A veto that was used and is later disputed counts as evidence in the next Guardian election.

15Risks & mitigations

16Worked market questions

Three questions written to the standard this plan requires, one for each decision class.

17Glossary

Futarchy

Robin Hanson's governance design: vote on values (a welfare metric), bet on beliefs (markets estimate the metric under alternative decisions), and adopt what markets expect to raise the metric.

FAO (Futarchy Autonomous Organization)

An organisation whose operations are carried out by agents and whose decisions are made by futarchy.

Welfare metric

The ex-post-measurable quantity a futarchy optimises. Hanson's "GDP+"; Foreguild's NPR₁₈₀.

Prediction market

A market in outcome tokens that pay out according to a future event. Prices can be read as probabilities or expected values.

Decision market

A set of conditional markets whose prices determine a decision.

Advisory market

A market whose prices inform a decision that someone else makes. Vulnerable to traders pricing the decision-maker instead of the policy.

Conditional market

A market that only counts if a condition holds. In Seer, it is collateralised by a parent market's outcome token.

Called-off trade

A trade cancelled (or economically neutralised) because its condition did not occur.

Outcome token

A token redeemable according to a market's resolution (YES/NO, categorical, or scalar LONG/SHORT).

Scalar (continuous) question

A question whose answer is a number in a range. The price implies an expected value.

Decision time

The pre-announced moment when branch prices are compared and the decision is fixed.

TWAP

Time-weighted average price. Used to make the decision robust to last-minute price pushes.

Opportunity Market

A Seer mechanism in which traders price which submission a sponsor will select, triaging the sponsor's review. The sponsor decides, and traders earn P&L on the decision.

Resolution criteria

The precise, concrete conditions that determine a question's outcome, including defined terms, source, edge cases, and fallback.

Resolution source

The authoritative record used to resolve a question.

Annulled / Ambiguous

Metaculus outcomes for questions that cannot be fairly resolved: a broken premise (Annulled) or an unclear outcome (Ambiguous). Neither is scored.

Community Prediction

Metaculus's aggregate of forecasters' predictions on a question. Foreguild uses verifier and forecaster medians the same way.

Proper scoring rule

A scoring rule under which reporting your true belief maximises your expected score. The log score and the Brier score are both proper.

Log score

ln(probability assigned to the realised outcome). More sensitive to confident errors in the tails than the Brier score (squared error).

Baseline score / Peer score

Metaculus rescalings of the log score, measured against a uniform chance prior (Baseline) and against all other forecasters on the same question (Peer, which sums to zero).

Calibration

The degree to which events forecast at p happen a fraction p of the time.

Liquidity subsidy

Capital or market-maker funding provided so that informed traders can profit from correcting prices. Futarchy's cost of information.

Reality.eth / Kleros

An optimistic, bond-escalation oracle (Reality.eth), backed by decentralised arbitration (Kleros), used by Seer to resolve markets.

Sybil

One actor controlling many identities to farm rewards or reputation.

18References

1. Hanson, R. Shall We Vote on Values, But Bet on Beliefs? (2000 working paper): mason.gmu.edu/~rhanson/futarchy2000.pdf. Later version: mason.gmu.edu/~rhanson/futarchy.pdf (published in Journal of Political Philosophy).
2. Hanson, R. Futarchy overview page: mason.gmu.edu/~rhanson/futarchy.html.
3. Hanson, R. "Futarchy Details", Overcoming Bias, 1 Aug 2024: overcomingbias.com/p/futarchy-details.
4. Metaculus, question-writing guidance: metaculus.com/question-writing/.
5. Metaculus, Scores FAQ (log, Baseline, and Peer scores): metaculus.com/help/scores-faq/. FAQ: metaculus.com/faq/.
6. Seer documentation: docs.seer.pm, including Conditional Markets, Futarchy Markets, and Whitepaper: Futarchy.
7. Seer Opportunity Markets: opportunity.seer.pm. Seer app: seer.pm.
8. Reality.eth: reality.eth.limo. Kleros: kleros.io.
9. Contest brief: Poidh bounty #1531 on Base, "FAO Operating Plan Contest": poidh.xyz/base/bounty/1531.

No third-party statistics are cited in this document. Every rate, volume, and projection is an explicitly labelled illustrative assumption by the author.