Hour audit: 2026-08-20T17

back to the files for this hour · how the archive is audited

Every number on this page is a stored measurement read from this hour's own records: its manifest.json, its witness statistics sidecar, and the corpus attestation. Nothing here is computed from the served bytes at the time you load the page, so no line below claims the bytes are correct. The digest and the one command that checks them are printed instead.

1 finding

• the hour is short of its sixty minutes. 58 of the 60 minutes carry events; 2 do not.

3 checks could not be evaluated for this hour, listed with the reason in the block below. A check we could not run is recorded as exactly that, and never counted as a pass.

The report

POLYMARKET ORDERBOOK ARCHIVE — HOUR AUDIT
hour 2026-08-20T17 · corpus v3 · verdict AUDIT PARTIAL (1 finding)

COVERAGE    minutes present 58/60 · first event 2026-08-20T17:00:00.000216+00:00 · last event 2026-08-20T17:59:59.978626+00:00
VOLUME      85,126,458 rows · 791,455,280 bytes (754.79 MiB) · 7 products, all holding rows
WITNESSES   not measured (this hour carries no per-machine row accounting)
INTEGRITY   sha256 2b214bb39880832d2956cf5e42375f9fd40dd671537d1a38bdab29950059dffc
            the corpus checksum ledger names this file with the same digest
SIDECARS    none: this hour's manifest has no v31 key. The v3.1 sidecars are optional, and hours before them have none
MARKETS     not measured (the per-hour market census runs on the capture host and its output is not published per hour)
HISTORY     layout union-event-sorted-v1
            re-merges not measured (the merge ledger records the merge that produced this file, not the history of attempts)
            supersessions not measured (supersession is recorded per corpus in the storage registry, not in the hour manifest)
CHECKS      4/8 pass, 1 finding, 3 not evaluated
              PASS    all products present
              PASS    row counts reconcile
              PASS    byte ranges tile the file
              SKIPPED the merge delta could not be checked
              SKIPPED which machines were expected is not published beside this hour
              PASS    the checksum ledger names this file
              FINDING the hour is short of its sixty minutes
              SKIPPED whether a boundary guard ran is not recorded
UPSTREAM    our incident register was checked against the venue's own status feed at 2026-09-23T20:35Z and names no event in this hour. That is a statement about what they published, not a check that every message they sent reached us.

The words in this report

Each line in the CHECKS block states what was found, not what was hoped for: a check that passes is named for the thing that is true, and a check that does not is renamed to the thing that is wrong. So FINDING never sits beside a sentence asserting the opposite of the one below it.

The report above is fixed-width text, so nothing in it can be clicked. The same terms are linked here instead. Its verdict is a statement about minutes and nothing else. The row it reports comes from this hour's own manifest and its witness-statistics sidecar, and the completeness line comes from the corpus attestation. The gap between the joined and merged counts is deduplication, not loss, and the boundary drops are the duplicates discarded at the seam with the hour before. Where the history line names a supersession, this hour was rebuilt to a better state after it was first published. The precision is the timestamp resolution these files carry.

Detail

The same records, per product. A product holding no rows renders its zero rather than being dropped: dropped and empty are indistinguishable to a reader, and only one of them is a defect. A cell reading not meas. is one nothing published, which is a third thing again and never a zero.

The class column is the merger's own verdict for that product, and it is keyed to CONTRIBUTION rather than presence. FULL means every machine expected to be capturing this hour delivered rows for that product; PARTIAL means some did and some did not; SINGLE-BY-DESIGN means only one machine was ever expected, and SINGLE-BY-DEFECT means several were and one delivered. So PARTIAL here is a statement about the fleet, not about the product being damaged or truncated: the rows that are present are whole.

PRODUCT                     rows          bytes   row groups  sha256
best_bid_ask           6,457,504     63,209,780         0-24  32eb1d3826f7096fd7a09f0994a95718001bec1c3b38f5460e4eee578752dbe1
book                   2,628,697    103,217,053        25-35  fb6f0c8fec065adc7b3141355bec2eb1d67a96eb77b115d2a058b161d432f442
last_trade_price          74,133      3,871,220        36-36  c2abfabc1052dadb1e684ca2f4480c7ce78a4c58d9bfad8ffee885764ed82f4d
market_resolved            1,501        116,551        37-37  b901a02443fda7c8e1851b33816ddf0f5db224a0b5234082d2953958f079c29d
new_market                 2,126        181,303        38-38  c63f1b0f2a0e814a555c379c14acd97a69b88765c82beaef5fba2da114c02d36
price_change          75,959,453    619,889,605       39-328  4db1671e5c5dcfa39386cfb606f961975a149496d5437557e1cb5ec2036d159c
tick_size_change           3,044        105,103      329-329  d5efc2d320c710b1823e6a800d453a17dd66e921a207cb478ab16e4c5f7a597c

EVENT MIX                  count    share
price_change          75,959,453   89.23%
best_bid_ask           6,457,504    7.59%
book                   2,628,697    3.09%
last_trade_price          74,133    0.09%
tick_size_change           3,044   <0.01%
new_market                 2,126   <0.01%
market_resolved            1,501   <0.01%

Captured and supplied are different numbers

Captured is how many rows each machine's own feed carried. Redundant capture means these OVERLAP, heavily and on purpose, so they do not partition the file and their sum is larger than it. Two machines each capturing nearly every event is the archive working, not double counting.

Supplied is how many rows of the merged file actually came from that machine. This one does partition: it sums to the file's row count exactly. But it is decided by the merge's tie-break, so a machine can capture almost everything and supply almost nothing. Read alone, either number tells a false story about the fleet, which is why both are printed and neither is called "the witness share".

What the skew columns mean

Arrival skew is the receipt timestamp minus the venue timestamp, in milliseconds. For six of the seven products the sidecar interprets it as transport latency, and marks that interpretation assumed: it rests on the venue timestamp being emit time, which we cannot verify from outside.

book is not latency and must not be read as latency. Its venue timestamp is the last time the book changed, so the sidecar interprets it as book staleness and warns in its own note that large values mean a quiet book rather than slow transport. Its maximum runs to weeks for that reason. This is why the table above gives a row per product per machine and never one figure for the hour: percentiles do not average, and these seven do not measure the same quantity.

Check it yourself

The digest above is a claim. This is how you refuse to take our word for it:

curl -O https://archive.pendulumflow.com/v3/2026-08-20/17/2026-08-20T17.parquet
sha256sum 2026-08-20T17.parquet
# compare with the sha256 line in the report above

A range digest certifies bytes. That those bytes are the product the manifest names comes from the manifest, which is why both are published and both are checkable.