A prediction market price is read as a probability about an event. It pays out on whatever the resolution mechanism certifies. Those are two different things. This is a public, timestamped log of how often they disagree — filed before resolution, scored both ways, and never edited.
No entries exist yet. This document is published before the first one deliberately, so that the rules governing what gets filed were fixed in public before anybody knew what the results would look like. If the selection rule were written afterwards, none of the numbers this record eventually produces would be worth reading.
When a market asks whether something will happen, two questions are hiding inside it. The first is whether the thing happens. The second is whether the mechanism that settles the contract decides that it happened. Most of the time those give the same answer, which is why the distinction goes unnoticed.
Sometimes they do not. A market can settle against the plain reading of its own written rule — because the rule was ambiguous, because the named source reported after the deadline, because the wording admitted an interpretation nobody intended, or because whoever adjudicated had a position in the outcome.
Nobody publishes a running measure of how often that happens, and the parties best placed to measure it are the ones being measured. This record does three things and nothing else:
Anyone can say a resolution was wrong after it lands. Saying so beforehand, with a probability attached, against a rule you have to publish first, is a claim that can fail in public. That is the only version worth keeping.
Not a resolver. We do not propose outcomes, stake bonds, vote in disputes, or operate an oracle. Several well-funded efforts are building resolution infrastructure and they are better placed to do it. This record measures the results; it does not compete to produce them.
Not a rating. There is no trust score, no letter grade, no composite index of any platform or mechanism. A composite score cannot be wrong, which is exactly why it is worthless as evidence.
Not about people. Entries concern markets, rule text, and mechanisms. No named trader, proposer, voter, or founder appears in an entry, even where their conduct is publicly visible and even where it is the interesting part of the story.
Not trading advice. A divergence finding is not a position. Nothing here says what anyone should do with money.
This is the section that decides whether the rest of the page is worth anything. Every accuracy claim in this industry founders on the same thing: the results are real, and the selection was not. Filing only the markets that later proved interesting produces a flawless record and tells the reader nothing.
So the rule is fixed in advance, published here, and applied without exception. It is deliberately mechanical: nothing in it requires a judgement call by us.
There is no discretion at any point. We do not read the rule text and decide whether it looks interesting before choosing to file — the five markets are determined by venue, date, and size alone. Most will show no divergence at all. Those entries are the majority and they are what makes a divergence rate mean anything.
Declining to file a market that the rule selects is a failure, not an editorial decision, and it is logged in the corrections log as one. Changes to this section happen by dated amendment published here, never by quietly filing a different set.
v0.1 of this section, published earlier today, read: a market was filed if it met all four of — listed on a covered venue; at least $250,000 in reported volume at filing; a resolution date within 45 days; and the rule text showing at least one ambiguity marker, being no named settling source, an undefined qualitative term, a compound condition, or a named source able to publish after the stated deadline.
Why it changed. Before filing any entry we tested the rule against live markets. Four were assessed and none qualified. The largest markets inside the window turned out to be carefully drafted from templates — one named its settling source, linked the official calendar, defined rounding for off-bracket outcomes, and specified what happens if the source publishes late. The loosely worded markets were novelty contracts carrying four-figure volume, which the floor excluded. On this venue, ambiguity and size run in opposite directions, so the v0.1 rule selected for the markets least likely to produce the thing this record exists to measure.
The deeper problem. The ambiguity criterion required us to judge whether wording was loose before deciding to file. That is the discretion this section exists to remove, and a reader would have been right to ask what we chose not to see. Ambiguity assessment now happens inside the entry, as the divergence thesis, where it can be scored and shown wrong — rather than in the selection, where it cannot.
The cost is stated plainly: a mechanical top-five rule is duller, and it makes a null result more likely. That is the correct direction for the trade to run.
v0.2 of criterion 03 read: Of the markets meeting 01 and 02, the five largest by reported volume at the moment of filing. All five are filed. Every week.
What changed. Two definitions were added. No criterion was added, removed, or reweighted. Whether the selection differs from what v0.2 would have produced cannot be stated, because v0.2 did not determine one — which is the defect.
Why. Preparing the first filing exposed that v0.2 left two words undefined at the venues where they must be applied. Polymarket groups several individually-settling contracts under a single event page — a rate-decision event, for example, is one contract per outcome bracket, each with its own rule text — and the venues display volume at more than one level. Left undefined, both words would have required a choice at filing time, which is exactly the discretion this section exists to remove. So: a market is the individually-resolving contract, the unit carrying its own rule text and settling on its own terms; volume is the USD figure the venue displays for that contract at the moment of filing, evidenced by a dated capture. A figure displayed only for a grouping of contracts ranks no single contract.
This amendment is published before the first selection is made under it, for the same reason the whole document was published before the first entry.
What changed. A third selection criterion was added, excluding contracts that resolve on the traded price, level, or yield of a financial instrument or index. The remaining criteria are unchanged and were renumbered.
Why. Capturing the venues for the first filing showed that the highest-volume contracts inside the 45-day window are hourly and daily asset-price brackets — a cryptocurrency’s price at 11am, an index level at the close, a Treasury yield today. Under v0.3 the mechanical top-five would have been dominated by them every week. But an entry’s rule-implied reading carries a mandatory probability on the observable the rule names, and when that observable is an asset price, the number is an asset-price forecast — which Section 06 of this document forbids, and has forbidden since v0.1. Section 03 as written selected what Section 06 prohibits. The exclusion was always implied by the hard exclusions; the selection rule simply failed to encode it, and the failure only became visible on contact with the venue. This is the third amendment this section has needed in two days, each found the same way. That is stated here rather than smoothed over, because a selection rule that survives its author’s imagination and fails at the venue is the expected order of events, and the record of the repairs is part of the record.
The judgement-call check. A criterion is only admissible in this section if applying it requires no discretion. The test for this one: does the contract’s written rule name, as its deciding observable, the traded price, level, or yield of a financial instrument or index? A Fed rate decision resolves on a policy announcement, not a traded price — included. CPI resolves on a published statistic — included. “Bitcoin above $65,000” resolves on a traded price — excluded. If a contract ever genuinely straddles this line, it will be filed rather than skipped, with the ambiguity stated in the entry, since filing is the non-discretionary default.
Amendments to this section are versioned, dated, and published here with the old text retained. Any entry filed under a previous version keeps that version stamped on it.
Every entry is filed before the market resolves. An entry filed afterwards is worthless and publishing one would discredit everything above it.
Both probabilities are mandatory. The gap between them is the entire contribution of this record; an entry carrying only one number is not an entry.
On resolution the entry is appended to, never edited. The original text stands exactly as filed.
Both probabilities are scored independently by Brier score. Reading rules well while predicting adjudication badly is a real and interesting result, and it will be published as one rather than averaged away.
Resolution is on the entry’s own terms. Where a generous reading would make us correct and a literal reading would not, the literal reading governs, and the resolution note says so. The Federation Record has already had to apply that rule against itself once, in FR-001, and this record inherits it.
Participation and standing convert to nothing — no tokens, no airdrop eligibility, no governance weight, no discounts — permanently. That rule is carried over unchanged from the Federation Record and is not subject to amendment.
Saying a log is append-only is easy; proving it is not. Every ten entries, or monthly — whichever comes first — the full record is exported as plain text and its fingerprint anchored to the Bitcoin blockchain via OpenTimestamps. Old snapshots are never removed, because the chain of dated exports is what proves nothing changed in between.
Before any snapshot is stamped, the regenerated export is diffed against the previous one. Prior entries must come back byte-identical; if they do not, that is investigated and disclosed before anything is published.
Stated now, while it is still cheap to say, so that it cannot be quietly dropped later.
If the record is full of divergences and every one of them was called correctly, be suspicious — ask what was excluded, and check it against the selection rule in Section 03. If it contains entries where we predicted a divergence and none occurred, still published, still scored, then it is doing what it was built for.