Glossary

The words this site uses about its own data, and what each of them does and does not claim. Terms are linked from their first appearance on every page, so you can arrive here from wherever the word confused you.

Several of these exist because a reader guessed wrong, and the guess was always in the same direction: that a figure means more than it does. So the definitions say what a term excludes as well as what it covers. Where a term has a stricter technical reading, the Precisely line carries it.

small

An hour holding under a quarter of the median hour's bytes for its corpus.

Precisely. A fact about volume, never a verdict about capture. Several small hours are checked complete for all sixty minutes and simply had little to record: 2026-08-19T04 sits at about 5% of median and is whole. A quiet market and a capture problem look the same from the outside, which is why the marker names its reason where one is known and says so plainly where none is.

witness

An independent machine recording the same live feed. An hour's witness letters name the machines that supplied rows to the merged hour, and contribution is counted by rows supplied rather than by being online. More witnesses mean more cross-checking, not more data.

Precisely. Counted from the machines that actually supplied rows, not from the list of machines that were meant to. The manifest's own witnesses field is that second list: it names the machines expected to capture the hour, so it counts a machine that was running and supplied nothing.

boundary drops

Duplicate events discarded at the seam between two adjacent hours during the merge.

Precisely. A small count is ordinary bookkeeping. A spike is investigated and annotated on the hour rather than smoothed away, because the seam is exactly where a merge defect would show first.

coverage ratio

Of the markets that traded in an hour, counted from on-chain fills, the share for which a corpus holds a book.

Precisely. The denominator is theirs, not ours. A ratio computed against our own market list would measure how consistent we are with ourselves, which is not a question anybody is asking.

dedup

Identical events reported by several witnesses collapsed to one row.

Precisely. The rows from every machine, added together, exceed the rows in the merged hour. That difference is deduplication, not data loss. Redundant capture is the point of running several machines, so a large difference on a busy hour is the design working.

audit pending

We have not looked yet.

Precisely. Not a synonym for complete and never a verdict. A section whose check has not run yet says this and nothing else. Otherwise the page would read as an all-clear that nobody measured.

verdict

An hour's completeness on minutes. Complete means all sixty minutes are present. Partial means some are not. Could not classify means the check ran and could not reach a verdict; those hours are counted on /audit.

Precisely. A partial hour states which minutes are missing, with the first and last timestamps actually present and a mechanism where one is known. Neither word says anything about whether the rows in the hour are correct.

supersession

An hour rebuilt to a better state, for instance when a witness that had been missing is healed in.

Precisely. The archive improves in place and the history is listed on the hour's own report, so a reader who downloaded the earlier version can see that it changed and why.

manifest

An hour's machine-readable record of what it contains and the digest of each part.

Precisely. Every number on an hour's audit page is read from its manifest, its witness-statistics sidecar or the corpus completeness record.

sidecar

A companion file published beside the data, such as witness statistics or a checksum ledger.

Precisely. Sidecars are derived from the corpus and must be measured. The incident register is written by hand, not derived from the corpus, so it is not published beside the data files.

attestation

The published per-hour completeness record the site renders its verdicts from.

Precisely. This is different from the coverage ratio on /audit, which counts traded markets against on-chain fills. The attestation covers minutes in our own hours. The coverage ratio covers all four corpora, six drawn hours each.

era

A dated change in how hours were produced. The eras are listed on /formats.

Precisely. Comparisons across an era boundary carry a line break rather than a smoothed join, because a curve drawn through a change of method looks like a trend when it is only a change of method.

precision

The timestamp resolution an hour's files carry, recorded per hour rather than assumed for the archive.

Precisely. Read it from the hour's record before comparing receive timestamps across hours; the published hours read microseconds where it is stated. The exchange's own timestamp is milliseconds throughout and is a different field from the receive timestamp; neither is interchangeable with the other.

snapshot grade

Hours that record book states without carrying every event between them.

Precisely. A statement about method, not about worth. The word snapshot here names the grade and not the book_snapshot event type, which is a different thing with a confusingly similar name.

A word that is not here

If this site uses a term about its own data that this page does not define, that is a defect rather than an oversight on your part. Every linked term has an entry. A term nobody linked can still slip through. Tell us and it gets an entry.