a manifesto for phil
Run the thought experiment honestly. A civilization that takes economics as seriously as we take physics has a trillion dollars to spend on public goods. What does it build?
Not foundations. Not application queues, biannual board meetings, or program officers. Nobody designing from first principles would invent the grant committee — a small group allocating other people's money by discussion, with no prices, no odds, no track records, and no consequences for being wrong. We have grant committees for the same reason we had scribes before printing: the better technology wasn't affordable yet.
So derive it. What would the sane civilization actually build?
(Readers visiting from dath ilan may skip ahead; sections 2–5 are your kindergarten curriculum.)
For private goods, the market is a complete loop: prices aggregate dispersed knowledge (Hayek), profit pulls talent and capital toward what people value, and loss prunes error without anyone having to admit anything. The loop runs by itself.
For public goods, every part of that loop is missing. There is no price that reveals what a public good is worth, so production is guesswork. There is no profit in producing one, so the most ambitious people build apps instead. There are no odds on whether an intervention works, so beliefs go untested. And there is no score for judgment, so influence over money flows to reputation and proximity instead of accuracy.
Public goods are not underproduced because people are selfish. They are underproduced because four markets are missing.
Above the four sits a clearinghouse. Funders who disagree about ends still have gains from trade (Ord, Moral Trade), and a mechanism like the S-process captures them mechanically: disagreement about values, cooperation about allocation.
This is the load-bearing hygiene of the whole design. Values are owned, not discoverable — no market can tell you what to care about. Beliefs are factual, and must be bet and scored. The grant committee's defining sin is laundering one through the other: "I value this" when the truth is "I predict this works," and back again, in the same sentence, with nobody keeping score.
The architecture should make the separation structural. Values enter as a funder's stated objectives, owned and signed. Everything downstream of "will it work" resolves against reality.
Mechanism design ran ahead of its information budget for seventy years. Lindahl prices need continuous preference elicitation. Outcome markets need cheap, credible resolution — did it work, and would it have happened anyway? Belief markets need thousands of well-posed questions written, refereed, and adjudicated. All of it was too expensive, so civilization fell back on the cheap substitute: trust inside small elites. Committees. Vibes.
That cost structure just collapsed. Evaluation, verification, question-writing, counterfactual analysis — the information work that made these markets unaffordable — is exactly the work AI makes cheap. The designs were never wrong; they were unaffordable, and now they aren't.
And the same technology is generating the largest pool of liquid altruistic capital in history — AI-lab equity, much of it already pledged. The wave and the toolkit are arriving together. That is not a coincidence. It is a deadline.
Not better grant reports. The system works on the day a talented twenty-five-year-old chooses to produce public goods instead of joining a startup — for financial reasons. Doing good becomes a career with prices, profit, and a compounding track record, instead of a vow of poverty with a committee at the end.
Own your values. Bet your beliefs. Pay for outcomes. Score the judges.
Phil is a portfolio of small, open, explorable implementations of the missing markets:
quadratic_funding — the demand market: broad support, not deep pockets, steers the pool.impact_hub — the outcome market: certificates, speculation, retroactive purchase of verified impact.calibration_games — the judgment market: resolved questions about real allocations, building the first track records.s_process — the values clearinghouse: many funders, one allocation, moral trade captured mechanically.recursive_regrants — the honest approximation: where stakes are small and verification dear, delegation through public trust chains routes money to local knowledge. Trust as credit, recorded so it can be priced later.grant_scorer, theory_of_change — the resolution layer: driving the cost of evaluation toward zero, which is what keeps every market above honest.Two commitments hold it together. First, the coordination layer must be a protocol, not a platform: prices, odds, and track records are themselves public goods, and if one operator owns them they will be enclosed and rented. Everything here is runnable, forkable, and open to any agent — human or AI — acting for any funder.
Second, honesty about the interim. Today these mechanisms still have operators, so every operator step snapshots to public, replayable artifacts that anyone can verify. That is scaffolding, not the cathedral. The end state is legitimacy by construction — mechanisms whose incentives are right under common knowledge, where there is nothing left to audit.
Money is about to stop being the bottleneck. Let's not let allocation become the excuse.