V3 = Military Grade Polymarket Orderbook Archive

Index of Polymarket orderbook captures

Military Grade - good enough to replay the orderbook: several machines, exact de-duplication, microsecond arrival times, coverage checked hour by hour. What that means.

Snapshot Grade - one recorder, millisecond timestamps that often collide, and no hour-by-hour check that the hour is whole. What that means.

8,245 files in all, counting LICENSE.txt. That is everything this domain serves.

Start with one hour

No account, no key. This is the most recent hour we have published:

curl -O https://dl.pendulumflow.com/v3/2026-10-10/01/2026-10-10T01.parquet

Check it against what we published, before you trust a byte of it:

# the sed strips the day/hour path so the name matches your flat download
curl -s https://archive.pendulumflow.com/v3/SHA256SUMS.txt   | grep 2026-10-10T01.parquet | sed 's|  .*/|  |' | sha256sum -c -

Then read it with DuckDB, a free command-line database (the beginner guide walks through installing it):

duckdb -c "SELECT * FROM '2026-10-10T01.parquet' LIMIT 5"

Each prefix has its own checksum file. /v3/SHA256SUMS.txt covers the hours we captured, and the mirrored eras have their own under /pmxt/v1/ and /pmxt/v2/. That hour also ships a manifest.json with a checksum for every event type inside the file, if you want to verify a part rather than the whole. The format notes explain the columns, and the audit says what we have checked and what we have not.

Data wanted, bridge the gaps

72 hours are missing between 2026-02-21T18 and 2026-10-10T01, in 3 separate stretches. If you hold Polymarket orderbook data for any missing hour, please get in touch. We will publish it here with credit. Part of an hour is worth sending.

68 of those have no known source, which is where help would matter most.

And we want better recordings of hours we already hold. Everything before our own capture is snapshot grade from a single source, and the audit has measured what that costs: v1 is missing about a third of the markets that traded, and v2 about an eighth. So we want your data even where this archive already serves the hour.

It does not have to have been recorded by several machines. That is how we reach military grade in our own capture, and it is not the price of contributing. A single recorder is worth having if it fixes any of these, and you can check them in your own columns:

• Receive times finer than a millisecond. The eras we mirror record milliseconds, which tie constantly, and events sharing a millisecond have no recoverable order.
• A per-event id or sequence number. With one, duplicates collapse exactly. Without one, duplicates can only be matched on their fields, which is a guess.
• Events kept in the order they arrived rather than written out in batches. Batching destroys the order inside each batch, and no later pass can recover it.
• A recorder that is not the one we already hold. Three of the AG6 hours turned out byte-identical to PMXT's, and a copy that could not have disagreed corroborates nothing.

For the hours listed above as absent we will take anything at all, at any grade, because nothing is what we have now. A second independent recording of an hour we already hold is still useful, because it is the only thing that can show what our capture missed. If you are not sure whether yours qualifies, tell us what its columns are and we will look.

https://x.com/PendulumFlow