Coverage report

This site no longer publishes a bare total of “crossings”. The unit is the event: one dated change of recorded affiliation, however many people it moved. A day on which several members left one party together is one event with several participants, not several events, and a person who crosses three times generates three events. The three units — event, participant, person — answer different questions and are not interchangeable.

A second distinction matters more than the first. The number of records the database holds is not the number of records the project stands behind. A record is publishable as verified only when its status is verified, its verification level is corroborated or official-and-independent, and it carries a publication date — a rule enforced by a database constraint rather than by editorial habit. Both figures are given below, and any citation should say which one it is quoting.

The project’s position is that these are two different acts and must never share a number. Leaving your caucus to sit as an independent is the legitimate response to disagreeing with the party you were elected under. Carrying the seat your voters gave to one party over to another, without an election in between, is the act this record exists to document. The first sentence is an opinion, it is the project’s, and it is argued on the methodology page. The second half of it — which of the two a given change was — is a question of fact about two stored labels, and the figures below answer it.

The unit below is the participation leg: one person, in one event. It is not the event count above, and the two are not interchangeable — a single event can hold ten legs of three different classes. Each figure carries the same population, so the acts can be compared honestly, and the buckets partition that population exactly rather than overlapping or leaving a remainder.

The comparison every editor asks for is with the Library of Parliament’s published figure. The two numbers are not in conflict; they are measured in different units, over different chambers, under different inclusion rules. Putting them side by side without converting first produces a false discrepancy, and the conversion is the credible part of the answer.

Three structural differences account for most of the gap, and none of them is a disagreement about the facts. Their figure counts people; ours counts events, and the person figure below is the one to compare. Their figure covers both chambers of Parliament; ours holds Senate affiliation events but publishes none of them, so a senator is present in the record and absent from every published total. And our group records that have not yet been resolved into named members hold people our person count cannot see, each record carrying its own declared estimate.

Publication status and verification level are two axes, not one scale. Status says how far a record has travelled through the editorial process; level says how well evidenced it is. A record can be well evidenced and unpublished, and an inherited lead can sit in the record for months without anyone having checked it against a primary source.

One trap in these two breakdowns is worth naming, because anyone querying the data directly will meet it. Filtering on the verification level alone admits records the project has retracted: a withdrawn record can still carry the verification level it had before it was withdrawn. Any query for publishable records must also exclude withdrawn ones, which is what the publication gate does.

Date precision is part of the fact, not a display preference. A record dated to a year is stored on the first of January of that year so that it can be sorted, and that stored day means nothing. Reading it as a real date — and in particular writing “1 January” into a story — is the single most likely way to misquote this database.

Facts here are stored as individual claims bound to citations, so “how well sourced is this database” has an arithmetic answer rather than an impression. The figures below are that answer. A source is held as a link and a locator, never as a copy of somebody else’s document, which is why link health is a live editorial concern and is published here.

Some events moved a group whose members the record has not yet resolved into named people. Those records carry an explicit estimate of how many people they involved, declared by the editorial record itself rather than projected here. Until they are resolved, the count of distinct people named in this database sits below the true number by roughly that estimate, and any comparison with a person-based figure elsewhere has to add it back.

A separate figure below is easy to mistake for a gap and is not one. Many participant legs have no destination party, because the member did not join a party: they went to the cross benches. There is no party row for “Independent” and there should not be one, so those legs are recorded with the affiliation label written at the time rather than with a party identifier.

These are structural limits, not backlogs. They will not close by adding more records, and a question that depends on one of them cannot be answered from this database at all — which is worth knowing before an hour is spent trying.

The outcome figures deserve particular care. Whether someone was re-elected, or was appointed to cabinet, is recorded for only a small minority of participant legs, so no rate can honestly be computed over the whole population. Where a share is quoted from this data it must be taken over the legs where the outcome is actually recorded, and published next to that number. Causation is never stored as a fact at all: an appointment dated beside a crossing is two dated facts, not a reward.

A bilingual site that quietly serves English text on its French pages is misleading a francophone reader, so the translation state of the record is measured and published here rather than left to be discovered. The French title column is never empty, which means an emptiness check would report full bilingual coverage; the honest measure is how many French titles are byte-identical to their English counterpart, and that figure is below.

Identical strings are not proof of an untranslated record — a person’s name and several party names are the same in both languages — so the figure is a review flag, not a verdict. Recorded translations also carry how much machine assistance produced them, and that breakdown is published too.

A single “last updated” date for a national dataset tells a provincial reporter nothing. What matters is the newest record this database holds for one jurisdiction, and how it was dated and verified. The table below gives exactly that, for every jurisdiction in the institutional table — including the ones where the answer is that we hold nothing at all.

Read the age column together with the precision column. A jurisdiction whose newest record is dated only to a year may in fact be more current than the date suggests, and a jurisdiction whose newest record is several years old is a place where this database should not be treated as a current source without checking.

Records change, and the changes are kept. Every revision is appended rather than overwritten, published corrections are listed in their own ledger, and machine proposals wait for a named human before they touch anything.