Audit: pmxt/v1

DownloadDocs

What this page is for: the problems with this corpus and the evidence behind them. The files are on the download page, and what the columns mean is in the format notes.

What is wrong with this data

Treat v1 as a partial sample of its era: useful for presence and shape, never for completeness claims.

• About a third of traded markets are missing. Against on-chain fills, v1 has a book for 65.00% of the markets that traded, across the six hours we picked in advance for the audit, ranging 32.24% to 99.94% hour by hour. Separately, on the hours v1 shares with v2, v1 lacks 41.93% of the markets v2 holds, and v2 lacks 2.28% of the markets v1 holds. Two methods with no shared input point the same way. the audit's check against on-chain fills, and the v1 to v2 overlap comparison, 2026-08-31
• Roughly half of v1's fills are phantom, and that is the venue's behaviour rather than a fault in this archive. 46% to 48% in every measured v1 hour. Polymarket's relayer reported fills that the on-chain contracts then rejected; traders called these ghost fills, and Polymarket's own exchange upgrade of 2026-04-28, which brought new contracts, a rewritten order book, and fees charged at match time instead of order time, was aimed squarely at that class of failure. PMXT recorded what the venue reported, which is what an archive is supposed to do. The phenomenon ends at 2026-06-19: markets from that date onward carry none. No market is entirely phantom: every traded market has at least one real fill. So this affects fill counts, not which markets appear. Polymarket's note on the upgrade. the audit's check against on-chain fills for the percentages, Polymarket's upgrade notes for the cause, 2026-08-31
• The real data ends earlier than the listing suggests. v1's last hour with data is 2026-04-15T08. The 21 hours listed after it, 2026-04-15T09 to 2026-04-16T05, are zero-row placeholder objects of 214 bytes each. They are listed because the files exist. They hold nothing. hour census against the published v1 SHA256SUMS.txt, with object sizes, 2026-08-31
• No transaction hashes, so individual fills cannot be matched to the chain. v1 records no trades and has no transaction_hash column, so individual fills cannot be matched to the chain. The coverage ratio above matches markets, not fills. read from the files, 2026-08-28

What each hour holds

This era is finished and nothing is still being written to it, so its contents can be stated once rather than watched. Every one of its 1,283 hourly files was measured: 30,264,647,169 rows. Files run from 2026-02-21T18 to 2026-04-16T05; data ends at 2026-04-15T08. The median file holds 23,914,475 rows, counting the 23 empty files; the largest holds 51,191,759.

23 hours hold no rows at all. Each is a 214-byte file: a parquet footer with no row groups, which downloads and opens and contains nothing. 21 of them are one unbroken run, 2026-04-15T09 to 2026-04-16T05, ending at the last hour of the era. Every hour of that run exists in V2 with between 52,297,097 and 67,127,329 rows, measured in the same pass, so the run is this capture winding down after V2 took over and those hours should be taken from V2. The other 2 (2026-02-24T14, 2026-02-24T15) fall before V2 begins: nothing here covers them and nothing here explains them.

What was measured, and what was not

• Rows: measured. The row count in each file's own parquet footer, read directly from the object, 1,283 of 1,283 files, none unreadable. Measured 2026-09-04T23:09Z.
• What we hold matches what was published: measured. All 4,118 mirrored objects across both eras are byte-identical to the source's own published checksums, and this site publishes a checksum for every one of them.
• Whether those rows cover everything that traded: not measured for the era as a whole. Six drawn hours were graded against on-chain fills (the table below); the rest have not been, and six hours is not a population rate. Grading every hour needs a per-hour comparison that has not been built. Until it is, this is not measured.
• Whether an outage explains a thin hour: not observable here. Our incident record reaches back to 2026-08-18T06, which is after this era ended, so no hour of it can be checked against a published incident. Those hours are unobserved, which is not the same as uneventful.

Row counts read from each file's parquet footer by ranged request, one pass, 2026-09-04T23:09Z. The full per-hour table is kept as rows-2026-09-04T230900Z.tsv, sha256 23b7730dfb45ba8e.

Measured against the chain

Every figure here is measured against on-chain fills for the same hour, by a program that reads the chain and never reads this archive. Six hours were drawn, and which six was written down before anything was measured. Six hours is a floor, not a sample large enough to describe a corpus, so nothing here is a population rate.

Our own capture scored 1.0000 on the same check. This corpus is at 0.6500, so roughly 35 in every 100 markets that demonstrably traded have no book here at all. A low coverage number means the market is missing, not that the data is imprecise. A book that was never recorded cannot be rebuilt from anything downstream. See all four side by side.

Pre-registration 5ae4e8765d601145, draw 02a94091dea9c11f, measured 2026-08-31.

Markets held, against markets traded

A corpus holds books for many markets that nobody traded in a given hour. Those books are what let you ask about a quiet market, not only a busy one. This counts both sides for each drawn hour.

This corpus has no single figure. Across the drawn hours it runs from 3.1x to 8.2x, and in date order the direction is plain, so a median of 3.9x would describe no hour it actually had. It is shown by hour, oldest first.

Our own capture holds about 52 books for every market that traded; this one holds about 4, roughly 13.2 times fewer. What it holds can still be trusted: this is a statement about how much of the market it watched, not about whether what it recorded is right. See all four side by side.

Breadth is not completeness. This number says nothing about whether the markets that did trade have books; the coverage ratio above measures that. The two are computed from different counts and rank the four corpora in the same order.