The final act of solving the puzzle is building the right fractal machine.

The market prices what is. Nobody is paid for pricing what isn't yet.

Every price is a statement about the present. Even a forward price is the present consensus about later. The market is exhaustively instrumented for the state it is already in, and completely uninstrumented for the state it has not reached.

Which is strange, because being early to that state is the single most valuable thing a person can do in an economy, and everyone already knows it. It is traded constantly — privately, in rooms, on reputation. Someone knew. Someone said it first. Someone was standing there before there was anywhere to stand. And there is no instrument for it. No score, no ledger, no way for a stranger to check, no way to carry it with you when you leave.

So it accrues to institutions instead of to people. A fund's returns belong to the fund. A firm's calls belong to the firm. The individual who actually saw it walks away with an anecdote. Judgment is the only asset in finance that its owner cannot take with them.

That is the whole opening. Not a technology gap — an instrument gap. And instrument gaps do not stay open, because the first credible instrument becomes the thing everybody cites, and everything cites one.

Four moves turn a hunch into an allocation right.

1. Make the claim capable of being scored

A claim enters as four things: the proposition, a stated confidence, a resolution condition written before the outcome, and a seal at the moment of filing so the date is not your word against anyone's. Anything missing one of the four is an opinion, and opinions cannot be priced. The seal is cryptographic and independent — the date has to survive you being unpopular.

2. Score it the way that isolates edge

This is where every "track record" product would go wrong. A proper score splits cleanly into two different things, and only one of them is worth money.

Reliability asks whether your stated confidence is honest — when you claimed near-certainty, were you near-certainly right. Resolution asks something else entirely: how far your calls sit from the base rate, in the right direction. That second one is edge. The first one is only the reason anyone should believe the second.

Three records make the same number of calls. Watch the rule pull them apart.

3. Measure edge against the prior, not against a market

A prediction market can only score propositions somebody made a market on — and making a market requires the proposition to already be legible. If a market exists, you were not early. Everything that carries real earliness is pre-market: this architecture wins, this founder is the one, this category is about to exist. No counterparty, no ticker, sometimes no vocabulary yet.

So consensus gets measured a different way. A language model trained to predict the most likely next thing is a compression of consensus — a mediocre generator and an excellent consensus meter. How surprising a claim was against that prior, on the day it was filed, is a continuous and reproducible measure of distance from the crowd. Edge becomes surprisal at filing × resolution at settlement, and it works in exactly the space markets cannot enter.

4. Pay in allocation rights, not in cash

A payout ends the relationship. An allocation right compounds it. Score converts into the ability to direct capital — your weight in where money goes next, growing as you keep being right and decaying when you stop.

This is the piece every allocation DAO missed. They route capital by token weight or by popularity, which measures capital already held, not judgment. Of course they underperformed a competent analyst; they were never measuring the thing that makes an analyst good. Route by scored judgment and the sentence people have been asserting for a decade finally becomes true: once that market is more efficient than the average allocator, the rest of the capital has nowhere better to go.

The allocation rule is itself a claim, so the machine contains itself.

Here is the turn. If being right routes capital, then the rule that decides how being right routes capital is the highest-value proposition in the system — and it is a claim like any other. It can be filed, sealed, resolved and scored by the identical machinery.

One rule, self-similar under change of scale. That is not a metaphor reached for after the fact — it is the literal definition, and it is why the thing is a fractal and why the machine that runs it is a factory rather than a fund. A fund allocates. A factory produces the allocators.

Numerai, unbounded.

Numerai proved the load-bearing part: strangers with no credentials, staking on their own predictions, aggregate into an allocator that trades real money. The mechanism works. It is also deliberately fenced on every axis — which is the opportunity, not a criticism.

Bounded, it is a hedge fund with a crowd attached. Unbounded, it is the pricing layer for human judgment, and the fund is one of the things that can be built on top of it.

Why a factory has to be underneath it

The index that scores earliness cannot be bootstrapped, because it requires elapsed time — you cannot buy a corpus of claims that have already resolved. It must be produced. A production line running many concurrent bets, each with a date and an outcome, is the only machine that manufactures resolved claims at industrial rate. That line already exists.

Persist Ventures production ledger and the CODECKS/0.1 genesis deck, both pulled 2026-08-08. Tracking begins Nov 2025 — roughly nine months of ledger, not a multi-year history. Nothing here is annualised and nothing here is a return.

What is honestly missing

The ledger measures input, not output — supply in fine detail and demand not at all. The instrument that collects it is built in-house, which is a real conflict: the party being measured should not own the only meter, and the standard has to be specified so that anyone else's meter can produce a conforming output. Sealed claims currently carry no confidence field, which means nothing sealed so far can be scored at all — that is the next change and it is one field wide. And a record where the author chooses what to file is a highlight reel, so coverage against a declared universe has to be part of the score or none of it counts.