Methodology

What counts as a floor crossing here, how the total is counted, how much weight each record bears, and where the evidence comes from.

This page is the specification the database is built to. It is published in full rather than summarised, because a compilation whose inclusion rules are not written down cannot be checked by anyone, reconciled with anyone else’s figure, or corrected on a point of principle.

It also contains an account of fabricated material this project once published and has removed. That account is here on purpose. The controls described further down exist because of specific failures named in it, and a reader deciding whether to trust this record is entitled to both halves of that story.

No count appears on this page. Every live figure — how many records exist, how many are verified, how well sourced they are — is measured on the coverage report.

Inclusion criteria and taxonomy

What counts as a floor crossing, why carrying a seat to another party and leaving a caucus to sit as an independent are counted apart, which contested cases are recorded and flagged rather than quietly dropped, and how the resulting figures can be reconciled with the Library of Parliament’s.

Why this document exists

A reference work is judged on whether its number can be audited, not on how large it is. This project published a total for years without publishing the rule that produced it. A reporter comparing our figure with the Library of Parliament’s could not tell whether the difference came from coverage, from definition, or from a counting error — so the figure was unusable in print.

Inclusion criteria have to answer two separate questions. First, what is one event? Second, which events belong in the headline total? This document answers both, class by class, and states for each class whether it is counted, merely recorded, or not yet representable at all.

The base definition

A floor crossing is a change in the recorded party or caucus affiliation of a person who holds a seat at the moment of the change, made without first resigning the seat and returning to voters under the new label.

Three conditions have to hold together, and each is separately checkable against a record:

The person held a seat on the date of the change. A candidate who switches parties between elections has no seat to cross with, and is out of scope.
The affiliation of record changed. Sitting as an Independent or as a non-affiliated member is an affiliation of record for this purpose — it is how the chamber lists the member, and it is how the Library of Parliament models the same thing.
No election intervened. If the label changed because the person was returned by voters under a different banner, that is an electoral fact, not a crossing.

We include and flag; we do not exclude in silence

The methodology this project published previously excluded expulsions where the member joined no other caucus, and said nothing at all about mergers, dissolutions, caucus renames or the Senate. Exclusion by silence is the worst of the options available: it lowers the total, gives no reader a way to raise it back, and guarantees a disagreement with every other compilation that nobody can locate.

The rule is now the opposite. Every class below is stored. Contested classes carry an explicit flag and, where appropriate, are held out of the headline total by default. The filtering happens in the open, at the reader’s choice, and it is reversible.

Two acts, kept apart

The base definition above covers both of the changes this record holds most of, and by itself distinguishes neither. A member can stop sitting with the caucus they were elected under in two ways: by joining a different caucus, or by joining none. This taxonomy previously treated both as feeding one total, and the database could not express the difference at all — the column that records how a leg counts had no value meaning “sat as an independent”, so a leg that ended on the crossbench either claimed the crossings total or stayed unclassified. Neither was true of it.

The distinction is now stored. Each participation carries a class saying how it counts: a crossing where both sides are parties and the parties differ; a departure to the crossbench where the leg ends outside every caucus; an arrival from the crossbench for the return; and separate values again for moves involving a parliamentary group, for renames, and for label changes at a general election. The class sits on the participation rather than on the event, because one event can hold more than one of them.

What the classification answers is a question of fact: did the seat go to another party, or to the crossbench? It is decided from the two affiliation labels the record already carried, by one rule applied in one place, so that a count and an audit of that count cannot come apart. A parliamentary group is deliberately neither a party nor the crossbench — it contests no election and holds no seat that voters awarded, so a move into or out of one carries nothing anywhere.

Why the project counts them apart, rather than merely being able to, is an opinion. It is published as one, set apart and attributed, at /methodology#project-position. It is not restated here, and it appears in no individual’s record.

Why this makes reconciliation with PARLINFO possible

The Library of Parliament states publicly that 307 MPs have changed political affiliation since 1867, of whom 158 moved from one party to another and 149 were Independents either before or after, with 80 in the last twenty-five years. A separate measurement of PARLINFO’s own "Changes in Political Affiliation" refiner, taken by this project on 31 August 2026, returns 438 flagged parliamentarians — 279 MPs and 159 senators — and 882 affiliation transitions reconstructed from their dated party-membership roles, 620 on the Commons side and 262 on the Senate side.

Those two figures already differ from each other, because one counts people and the other counts transitions. Ours differs from both again, for four reasons that are each stated and each reversible: we count events rather than people or role-periods; we cover provincial and territorial legislatures, which PARLINFO does not; we hold Senate affiliation events but have published none of them, so any figure drawn from our published record is a Commons and Assembly figure; and we record a multi-member move as one event rather than as one transition per member.

Reconciliation does not mean agreement. It means that the gap between two totals can be decomposed into named, checkable differences. Because every class carries a flag and every participant carries a person, a reader can rebuild the Library’s person-count, or its transition-count, from our data and see exactly where the remainder sits.

What is outside the record entirely

These are not contested classes held back by a flag. They fail the base definition, and no amount of filtering brings them in:

A candidate or former member changing party while holding no seat. It may be recorded as context on the person; it is never an event.
A change of cabinet portfolio, committee assignment or parliamentary office with no change of affiliation.
A leadership contest, unless a caucus affiliation actually changed as a result.
A reported intention to cross that no official record ever confirmed. These are held as research candidates and are not published as events, however widely they were reported at the time.

Questions this specification has not settled

Three cases remain genuinely open, and are listed here rather than resolved quietly in code:

Whether a party succeeding itself under a new name can be told from a caucus defecting en masse. The two have exactly the same shape in the data and opposite answers, and nothing the database holds separates them: the table of relationships between parties is empty, no event is scoped as institutional, and no party’s founding or dissolution date lines up with a leg. Every group of that shape is counted as a crossing, because that is what its labels say, and every one of them is listed for review so that the size of the unmade ruling travels with the figure.
Whether wartime coalition labels — the 1917 Unionist and Liberal-Unionist designations in particular — are mid-term crossings, election labels, or both at once for different members of the same cohort. The database currently holds this cohort as a single unresolved group lead.
Whether the items that the principal public aggregate of Canadian floor crossings relegates to an "other changes" section — chiefly mergers and pre-election party changes — should appear in a comparison total or only in the underlying data.

Text last checked against the database: 2026-09-01

The classes of event

1. Voluntary party-to-party crossing

Counted in the floor-crossing figureRepresented in the record

A sitting member leaves one caucus by their own decision and is received into another, while the legislature stands and without an intervening election.

Inclusion rule: Include when a record shows both the departure and the arrival on one date, and shows that the member held the seat on it. Where a period sitting as an Independent separates the two, this class does not apply: record two events, which fall under classes 2 and 4 respectively — the dates are different, each is separately verifiable, and the seat spent the interval attached to no party at all.
Counting rule: One event, with one participant per person moving. Several members arriving on the same day under one announcement form a single group event. This is the only class the floor-crossing figure is drawn from: a leg counts toward it when both of its sides are parties and they are different parties, and the participation is stored as `crossing`. Where a period sitting as an Independent separates a departure from an arrival, the two legs take two different classes and only the arrival is party-to-party — so an episode of that shape contributes nothing to this figure at either end. It is published as its own figure instead.
Stored as: changed_affiliationjoined_party
Counted as: crossing

Examples

Belinda Stronach joins the Liberal Party and Cabinet, 17 May 2005
In the database, verified · CTF-E-000072
Day-precision, corroborated. The clearest federal case of a same-day crossing and cabinet appointment.
David Emerson joins the Conservative Party and Cabinet, 6 February 2006
In the database, verified · CTF-E-000073
Day-precision, corroborated. Crossed within weeks of being returned under the previous label.
Chris d’Entremont joins the Liberal Party of Canada, 4 November 2025
In the database, verified · CTF-E-000090
Day-precision, corroborated. One of a run of federal crossings in the same direction across 2025 and 2026.

2. Voluntary departure to sit as an Independent

Counted in the sat-as-an-independent figureRepresented in the record

A sitting member leaves a caucus by their own decision and sits without party affiliation.

Inclusion rule: Include whenever the chamber’s own membership record changes to Independent or non-affiliated. A leave of absence, an illness, or a suspension that leaves the recorded affiliation untouched is not this class and is not an event.
Counting rule: One event, counted toward the figure for members who left a caucus and sat as an independent — never toward the floor-crossing figure, and never added to it. No party gained the seat, so there is nothing for a crossings total to record. The two figures are published side by side and answer different questions; the reasoning is at /methodology#project-position and is an opinion, plainly marked as one. If the member later joins another party or returns to the original caucus, that is a further event with its own date under class 4, not a revision of this one.
Stored as: left_to_independentcaucus_departure
Counted as: sat_as_independent

Examples

Alexandre Boulerice sits as an Independent, 27 April 2026
In the database, verified · CTF-E-000274
Day-precision, corroborated. Absent from the legacy public dataset entirely, which is why the deployed site showed a staler record than the database held.
Youri Chassin leaves the Coalition Avenir Québec caucus, 12 September 2024
In the database, verified · CTF-E-000283
Day-precision, at the highest verification level: an official record and an independent account agree.

3. Expulsion from a caucus

ContestedCounted in the sat-as-an-independent figureRepresented in the record

A caucus removes a member. The member sits as an Independent because the caucus acted, not because they chose to.

Inclusion rule: Include, and record the participant as involuntary. The methodology this project published previously excluded expulsions where the member joined no other caucus; that exclusion is withdrawn. The affiliation of record changed, the chamber published the change, and the Library of Parliament counts it.
Counting rule: One event, counted toward the same figure as class 2 — members who left a caucus and sat as an independent — and flagged involuntary so that a reader who wants only voluntary departures can filter it out and say so in print. The stored class describes where the member sat afterwards and not why, which is the only thing it could honestly describe: voluntariness is recorded separately and reads unknown on the large majority of legs, so the two cannot share a field. Expulsion is not a choice, and nothing in the classification implies it was one.
Stored as: expelledsuspended
Counted as: sat_as_independent

Examples

Marwah Rizqy is expelled from the Quebec Liberal caucus, 2 December 2025
In the database, verified · CTF-E-000290
Day-precision, official and independent sources agree.
Elenore Sturko is removed from the Conservative caucus, 22 September 2025
In the database, verified · CTF-E-000294
Day-precision, official and independent. The same member also appears at CTF-E-000133, an earlier affiliation change held at year precision with voluntariness unrecorded — which is why voluntariness is a property of the participation and not of the person.

4. Return to a caucus after departure, suspension or expulsion

ContestedCounted in neither figureRepresented in the record

A member who has been sitting as an Independent, whether by choice or by removal, is readmitted to the caucus they left.

Inclusion rule: Include as a separate event with its own date. Some compilations treat a departure and a return as one episode; we do not, because the return is a change in the affiliation of record, it is frequently the newsworthy half, and collapsing the pair destroys the interval a reader needs in order to say how long the member sat alone. That interval is the point. A member who goes to the crossbench, faces a general election and then sits under a new banner has done something the voters were asked about; a member who joins another caucus without an election in between has not. The two are the same shape leg by leg, so the record derives the answer per pair, from the dates, rather than storing it — and reports it as unknown wherever the election series is too thin to assert that no election happened.
Counting rule: One event, counted toward neither published figure, and published as a third figure of its own. Adding it to the crossings would count one person twice for a single episode, since the outbound leg is already counted as a departure to the crossbench; dropping it would hide the most interesting cases in the record. A departure and a return are two events by one person: the event total rises by two, the distinct-person total by one. This is the single most common way a floor-crossing count is misread, and the reason a figure must always name its unit.
Stored as: return_to_party
Counted as: joined_party_from_independent

Examples

Simon-Pierre Savard-Tremblay: Independent on 28 May 2026, back with the Bloc Québécois on 19 June 2026
In the database, verified · CTF-E-000276
The return, paired with the departure at CTF-E-000275. Both day-precision and corroborated. One person, two events, twenty-two days apart.
Lionel Carmant: leaves the Coalition Avenir Québec caucus on 30 October 2025, returns on 1 April 2026
In the database, verified · CTF-E-000287
The return, paired with the departure at CTF-E-000286. Both at the highest verification level.

5. Joining or founding a new party while sitting

Counted in the floor-crossing figurePartly represented

One or more sitting members leave their caucus for a party that does not yet exist, or that exists but holds no seats until they arrive.

Inclusion rule: Include. The destination need not be a registered party on the day of the move; a named caucus that the chamber recognises is enough. Where the party is registered later, the registration is stored as a dated fact about the party and never substituted for the date of the event.
Counting rule: One event however many members move together, with one participant per member. The founding of the party is a fact about the party; the members’ departures are the event.
Stored as: party_formationjoined_party
Counted as: crossing

Examples

Thirteen Social Credit MPs form the Ralliement des créditistes, 1963
In the database as an unverified lead · CTF-E-000158
Held at year precision as an unresolved group lead: the thirteen members are named in the title but not yet enumerated as participants.
Around fourteen Alliance MPs form the Democratic Representative Caucus, 2001
In the database as an unverified lead · CTF-E-000223
Year precision, unresolved. The approximation in the title is itself the finding: the legacy record never established which members and on what dates.
Two sitting Alberta MLAs relaunch a provincial Progressive Conservative party, July 2025
Documented in the research file, not yet in the database
Documented in the project research file, following one departure to Independent and one expulsion earlier in 2025. Not yet written to the database; listed here so the gap is visible rather than implied.

6. Party dissolution or merger absorption

ContestedCounted in the floor-crossing figureNot yet represented

The affiliation of record changes because the party ceases to exist, merges, or is absorbed — not because a member went looking for a new home.

Inclusion rule: Include, flagged. A member’s label changed on a stated date, in a chamber record, without an election. Omitting these is the single largest cause of silent disagreement with PARLINFO, which records each leg of a merger for each affected member.
Counting rule: One event, with one participant per member whose label changed. When a whole caucus moves in one day, the event total rises by one and the participant total by the size of the caucus. The caucus size must never be added to a crossings total. A party succeeding itself under a new name and a caucus defecting en masse have exactly the same shape — everyone on one event moving from the same party to the same party — and the record currently holds nothing that separates them. Legs of that shape are counted as crossings, because that is what their labels say, and the number of them resting on the unmade ruling is published beside the figure rather than left as a footnote.
Stored as: party_dissolutionmerger
Counted as: crossinglabel_rename

Examples

Seven BC United MLAs, 2024
In the database as an unverified lead · CTF-E-000088
Year precision, unresolved group lead, and the French title is still the English one. The underlying episode ran across at least 2023 and 2024 with members moving on separate dates, which a single year-stamped row cannot express.
The Canadian Alliance succeeding the Reform Party, 27 March 2000
Documented in the research file, not yet in the database
PARLINFO records twenty-seven same-day transitions for this succession. Exactly one event with twenty-seven participants under our rule; twenty-seven rows under theirs. Not yet in the database.
The dissolution of New Brunswick’s People’s Alliance in 2022, after which its two sitting members joined the Progressive Conservative caucus
Documented in the research file, not yet in the database
Documented in the project research file. The provincial case that shows why dissolution needs its own class: neither member sought the move.

7. Caucus or label rename

ContestedCounted in neither figurePartly represented

The name under which a member sits changes without the member moving: a caucus renames itself, a party rebrands, or a records system adopts a new label for the same group.

Inclusion rule: Record every leg, flagged as a rename, and hold it out of the headline total by default. A rename is not a decision by the member, and counting it as a crossing inflates every historical total. It has to be stored all the same, because PARLINFO stores it and because a reader reconstructing a member’s affiliation timeline needs the leg for the timeline to be continuous.
Counting rule: Recorded, counted toward neither published figure, and reachable through a filter. A leg is a rename when both of its sides resolve to the same body, however differently the two labels read; where both sides are outside every caucus it is stored separately again, as a change of designation for a member who was already sitting alone. Where a rename is bundled into an episode that also contains a real departure, the legs are split and only the departure is counted.
Stored as: otherchanged_affiliation
Counted as: label_renamecrossbench_relabelling

Examples

Seven Bloc Québécois MPs, the Québec debout episode, 2018
In the database as an unverified lead · CTF-E-000220
Year precision, unresolved. The textbook mixed case: a genuine departure, then successive renamings of the breakaway caucus, then a return. PARLINFO records the legs separately for each of the seven members. Our single year-stamped row cannot yet distinguish them, which is why it stays a lead.
Historical relabelling inside a single party lineage, such as a records system distinguishing a modern Conservative label from a Conservative label used before 1942
Documented in the research file, not yet in the database
Observed in the PARLINFO extract taken by this project. These rows are the main reason a raw transition count from that source overstates a journalist-grade event count.

8. Senate group change

ContestedCounted in neither figurePartly represented

A senator changes parliamentary group. Since the 2016 reform the Senate is organised around groups rather than parties, so what a senator sits with is frequently not a party at all.

Inclusion rule: In scope by the base definition — the affiliation of record changes and no election intervenes — and now held in the record. The database has an upper chamber, the Senate of Canada, and it carries Senate affiliation events. None of them has been published. Every one is a pending review, none has passed the publication gate that governs every other class in this document, and they are therefore unreviewed leads rather than a category held back for being Senate. The live split between what is held and what is published is at /coverage and GET /api/v2/coverage, and is deliberately not restated here.
Counting rule: Nothing on the Senate side counts toward either published figure today, because nothing on the Senate side is published. Two things keep the class distinct even once it is. Senate groups are not parties: the chamber organises around groups, so what a senator sits with often has no party behind it — a group contests no election and holds no seat that voters awarded, so moving into one carries nothing anywhere and moving out of one is not a withdrawal to sit alone. And a newly appointed senator moving out of non-affiliated status is institutional onboarding rather than a change of allegiance: it is the first affiliation the record holds for someone appointed to the seat, not a decision to leave one allegiance for another. The distinction is now stored rather than implied. `event_participants.affiliation_change_class` carries `appointment_onboarding` and `chamber_group_change` as separate values, set per participation because one event can hold both — a single day in December 2016 moved a group of senators of whom some were being onboarded and the rest were making a real change of group. A group move in an elected chamber is a different question again, and every such leg is listed for editorial review rather than settled by the Senate reasoning: a breakaway group of sitting members is closer to the act this site records than a Senate group change is.
Stored as: otherchanged_affiliationjoined_partyleft_to_independent
Counted as: appointment_onboardingchamber_group_change

Examples

Seventeen senators recorded as moving from Non-affiliated into the Independent Senators Group, 2 December 2016
In the database as an unverified lead · CTF-E-PLI-20161202-57F4CECD
Held, unpublished, and the reason the onboarding question cannot be settled at the level of an event. Some of these senators had been appointed shortly before, and this is the first affiliation the record holds for them. Others had been sitting for years and had reached non-affiliated status through a recorded departure, so joining the group is a real change of allegiance for them. One event, one date, two classes — which is why the classification has to sit on the participant and not on the event.
The measured scale of the gap: 159 senators flagged in PARLINFO and 262 senator transitions, 65 of them since the start of 2023
Outside the coverage of this database
Measured against the Library of Parliament’s own interface on 31 August 2026. The gap is no longer structural: the schema expresses a senator and the record holds Senate events. It is now an editorial gap — none of what we hold has been reviewed to the publication standard, so the published Senate figure remains zero while the held figure does not.

9. Election-time relabelling without a mid-term act

ContestedCounted in neither figureRepresented in the record

A person’s party label differs before and after an election, but nothing happened between elections: they sought and won nomination under a different party, or an election-period label differs from the label of record.

Inclusion rule: Exclude from the count. This is the one class the base definition rules out on its face, because the person went back to voters under the new label. Where an inherited record turns out to be one of these, it is withdrawn rather than deleted, retitled so that the withdrawal is legible, and linked to the verified in-term event if one exists.
Counting rule: Not counted. Retained as a withdrawn record at its original identifier, so that anyone who cited it before the correction can see what happened to it and why. This is the one class whose stored labels can read party-to-a-different-party while the ruling is that it counts toward nothing, so a reader must not infer the count from the two labels alone: the labels say what changed, and the class says whether the voters were asked.
Stored as: other
Counted as: election_relabelling

Examples

Withdrawn: pre-election Gilles Roch party change, 1986
In the database, withdrawn · CTF-E-000167
The worked example of the whole policy. The inherited 1986 entry was a pre-election party change, not an in-term crossing. It is withdrawn, its title says so, it is linked to the verified in-term event of 8 September 1988 at CTF-E-000277, and the reasoning is published in the corrections ledger as CTF-COR-OUTCOME-000167.

10. Jurisdiction-specific rules

Counted in neither figureNot yet represented

Not a class of event but a class of rule: local law or caucus procedure that changes what a crossing can be, or what evidence a crossing leaves behind, in one jurisdiction.

Inclusion rule: A local rule changes the meaning of an event, never whether it is recorded. Two rules bear directly on classification. Manitoba amended its Legislative Assembly Act in 2006 to require a member who leaves the caucus of the party under which they were elected to sit as an Independent for the remainder of the term, which makes an in-term party-to-party crossing structurally impossible there. At the federal level, the caucus framework adopted in 2015 gives a caucus a recorded vote on expelling a member, which changes the evidence an expulsion leaves behind.
Counting rule: No effect on counting. A jurisdiction with a restrictive rule will show few or no party-to-party events; that is a fact about the law and has to be labelled as such, or it will be read as missing data. Manitoba’s holdings are correspondingly thin, and one of the few records there is the withdrawn 1986 entry described above. The per-jurisdiction figures are published at /coverage rather than written into this page.
Stored as

Examples

Manitoba’s 2006 restriction, and the federal caucus-expulsion framework of 2015
Documented in the research file, not yet in the database
Both rules are recorded in the project research file. Neither is yet stored in the database with a citation to the statute or standing order, so no record relies on them and they are presented here as findings under review rather than as settled law.

Counting rules

The difference between an event, a participant and a person; why carrying a seat to another party and leaving a caucus to sit as an independent are counted as two figures and never one; why a caucus that moves together is one event and not many crossings; and how to compare our figures with the Library of Parliament’s or with a public encyclopedia’s.

What the headline number counts

The headline number is a count of events: distinct, dated changes of recorded affiliation by people who held a seat at the time, in the federal House of Commons and in provincial and territorial legislatures, excluding election-time relabelling and excluding caucus renames. The Senate is in the record but not in this number: the record holds Senate affiliation events and has published none of them, so they fail the publication gate rather than the definition.

It is not a count of politicians, and it is not a count of party-to-party moves. A person who leaves a caucus and later returns contributes two events and one person. A caucus of seven that moves on a single day contributes one event and seven participants. Nor is it a count of floor crossings in the narrow sense: an event total includes departures to the crossbench alongside moves between parties, because both are dated changes of recorded affiliation. The narrow figures live at the participation level and are set out in the next section.

Two figures, never one

Beneath the event total, every participation is counted by what it did with a seat, and the results are published as separate figures that are never added together. A leg whose two sides are both parties, and different parties, counts toward the floor-crossing figure. A leg that begins at a party and ends outside every caucus counts toward its own figure, for members who left a caucus and sat as an independent. The two answer different questions and a single number would answer neither.

The return leg — a member leaving the crossbench for a party — counts toward neither, and is published as a third figure of its own. Adding it to the crossings would count one person twice for a single episode, because the outbound leg of that episode is already in the crossbench figure. Suppressing it would hide the cases where somebody sat alone briefly and then joined another caucus with no election in between, which are among the most consequential things the record holds. So it is neither added nor hidden: it is its own number, published beside them.

Everything else is a published figure too, rather than a silence: moves into, out of or between parliamentary groups; the first affiliation of record for someone appointed to a seat rather than elected to one; renames; changes of designation for a member already sitting alone; label changes at a general election; legs suppressed as duplicates of a superseding record; and legs on which no ruling has yet been made. The categories partition the record exactly — every participation is in one and in only one — so the figures audit each other, and adding all of them together yields the total number of participations and nothing more interesting than that.

The unruled remainder travels with them and is not zero. It is the size of a decision nobody has taken yet, and a crossings figure published without it presents itself as more settled than it is. The same goes for two named caveats: how much of the crossings figure comes from the appointed chamber, and how much of it comes from groups of members making one move on one day, where the record cannot yet tell a party renaming itself from a caucus defecting together.

Group events: one event, many participants

When several members change affiliation together — a caucus splitting, a party dissolving, a bloc of members founding something new — the record holds one event and one participant row per member. The event carries the date, the class and the sources. The participant rows carry each person’s own origin party, destination party, riding, voluntariness and effective date, because those can differ from member to member even inside one announcement.

This is the rule that most often breaks in practice, and it breaks in two opposite directions. Adding the caucus size to the event total double-counts a single decision. Treating the group as one crossing by one person erases six or twelve or twenty-seven people from the record. Both errors have appeared in published work about Canadian floor crossings.

Some group records inherited by this project name a number of members in their title but have no participant rows behind them — "around fourteen Alliance MPs", "seven BC United MLAs". These are unresolved group leads. Each data release publishes how many there are, in the quality metrics recorded on the release itself, precisely so that the participant total can be read with the right amount of scepticism.

One episode is not one event

A departure and a later return are two events. A departure to Independent followed weeks later by arrival in another party is two events. A dissolution followed by absorption into another caucus is two events if the chamber recorded two changes, and one if it recorded one. The test is always the record, never the narrative: if the affiliation of record changed twice, there are two events, whatever a single news story called it.

Splitting episodes this way is what makes intervals measurable. It is the reason the record can answer how long a member sat as an Independent before returning, or how many days separated a crossing from an appointment — questions a collapsed single-row episode can never answer.

Where the authoritative totals live

Totals are not properties of a web page. Each data release records its own event count, participant count, person count and source count, alongside a set of quality metrics: how many events are verified, how many group records remain unresolved, how many leads carry at least one cited source, how many citations point at nothing, how many editorial reviews have been approved, and how many source links have been checked.

Cite a release, not a page. A release has an immutable version string, and its numbers do not move after publication. Anything quoted from a live page is a snapshot of a moving target, which is exactly the problem this project exists to solve.

Text last checked against the database: 2026-09-01

The three units

Event

affiliation_events

One dated change of recorded affiliation, however many people it moved. Carries the date and its precision, the class, the jurisdiction, the publication status, the verification level and the sources.

Use it for: How often crossings happen; how they cluster in time; era and jurisdiction comparisons; the headline total.
Do not use it for: How many politicians have crossed. One person can generate several events, and one event can involve many people.

Participant

event_participants

One person’s involvement in one event. Carries that person’s origin and destination party, the affiliation labels as written at the time, the riding, whether the move was voluntary, whether it was toward or away from government, the change in seat balance and that person’s own effective date.

Use it for: Seat arithmetic; direction of travel between parties; party-level gains and losses; anything that has to be true of an individual rather than of an announcement.
Do not use it for: The headline count of crossings, and not as a person count either: a person who crosses three times has three participant rows.

Person

persons

One human being with a stable identifier, linked to every participant row they appear in, across chambers and across decades.

Use it for: The comparison with the Library of Parliament’s published figure, which is a person count; "how many MPs have ever crossed"; repeat crossers.
Do not use it for: Frequency over time. People are not dated; events are.

Where the seat went

Read from the two affiliation labels this record carries, by the same rule the database applies. It states the direction of the move, not the editorial ruling on how the leg counts.

Where the seat went
Controlled vocabulary	Definition	What you may rely on	How it counts
The seat was carried to another party`to_another_party`	Both sides of the change are parties, and they are different parties. The member sat with one caucus and then sat with another, without the seat being returned to voters in between.	That a party gained a seat it did not win. It says nothing about why: the record separates where a member went from whether they chose to go.	The direction the floor-crossing total is drawn from, but not the total itself. A small number of legs whose labels read party-to-a-different-party are ruled election-time relabelling, and count toward nothing.
The member left a caucus and sat as an independent`to_the_crossbench`	The change begins at a party and ends outside every caucus. No party gained the seat. A designation such as “Independent Liberal” or “(Not in Caucus)” is on this side of the line: it describes an alignment, not a caucus membership.	Where the member sat afterwards, and only that. Whether they walked out or were expelled is `voluntary_status`, which is recorded separately and reads unknown on most legs; /coverage has the live share.	Counted as its own figure and never added to the floor-crossing total. They are different acts, and a single number that merged them could not answer either question.
The member left the crossbench for a party`from_the_crossbench`	The return leg: a member sitting outside every caucus joins one. Sometimes the caucus they had left, sometimes another.	That the member ended in a caucus. Whether a general election fell between the departure and this return is the question that separates two very different sequences, and it is answered per leg by `independent_return_intervals`, not by this badge.	Counted toward neither total. The outbound leg of the same episode is already counted as a departure to the crossbench, and adding this one would count a single person twice for one episode.
Independent on both sides of the change`within_the_crossbench`	The member sat outside every caucus before and after. The designation changed and the member did not.	That no caucus was joined or left. Most of these are changes of wording in the chamber’s own membership record.	Counted toward neither total.
A parliamentary group is on one side of the change or on both`parliamentary_group`	A parliamentary group is not a party and not the crossbench. It contests no election and holds no seat that voters awarded — the Senate has organised around groups rather than parties since 2016 — so moving into one carries no seat anywhere, and moving out of one is not a withdrawal to sit alone.	That the change involved a group rather than a party. Most such legs are in the Senate; the ones in an elected chamber are listed for review, because a breakaway group of sitting members is a closer thing to a crossing than a Senate group change is.	Counted toward neither total.
Both sides are the same body under two names`same_body`	The two labels differ but resolve to one party or one group. The name under which the member sat changed; the member did not move.	That nothing was carried anywhere. The leg is kept so that a member’s affiliation timeline stays continuous, which is why it is visible rather than dropped.	Counted toward neither total.
One side names something the record cannot yet resolve`unknown`	A label on one side of this leg is not mapped to a party row and is not in the enumerated lists of crossbench designations and unmapped party names. Nothing is guessed from its spelling.	Nothing about the direction. This is a gap in the record, published as a gap: a label nobody has mapped is reported rather than absorbed into whichever bucket a pattern match happened to reach.	Counted toward neither total, and reported as an unmade ruling rather than as a zero.

The project’s view — not a finding

Stated by Crossing the Floor. Nothing below is derived from the record, and no record depends on it.

Everything else on this page can be checked against a source. This cannot: it is an argument, and a reader is free to reject it and keep every figure on the site.

Why this record separates two acts

The project holds that leaving a caucus to sit as an independent and carrying a seat into another party without an election are different acts, and counts them apart. This section states that view, marks it as a view, and sets out what the record cannot tell you about it.

Whose view this is, and why it is written down

Everything else in this methodology is a specification. This section is not. It is an opinion, held by the project, about what the two acts recorded here mean — set apart, headed as an opinion and attributed, so that no reader mistakes it for a finding and so that a reader who rejects it can still use every number on the site.

It is published rather than left implicit for the same reason the inclusion criteria are published. The project makes a choice here — to separate two kinds of change that most compilations add together — and a choice whose reason a reader cannot see is indistinguishable from an error. A reference work called Crossing the Floor that claimed to have no view about crossing the floor would be less honest than one that states its view plainly and keeps it out of the record.

What the project holds

Disagreeing with your party is a legitimate reason to leave its caucus. A member who does so and sits as an independent takes the disagreement to the place where it costs them something, and leaves the electorate’s decision intact: the seat stays where the voters put it, unattached, until the next election — at which point the member may seek a mandate under whatever banner they choose. That is the response this project regards as the proper one, and the record holds a great many of them.

Carrying that seat into another party’s caucus, without an election, is a different act. The seat was awarded to a person standing under a label, and moving it to a different label transfers something the voters were never asked about — frequently in mid-term, sometimes into a governing majority, and always without the constituency having any say until the next general election. That act is what this record exists to document, and it is what the site is named after.

This is not a novel standard, and the project did not invent it. One Canadian legislature has already written it into law: Manitoba amended its Legislative Assembly Act in 2006 to require a member who leaves the caucus of the party under which they were elected to sit as an independent for the remainder of the term, which makes a mid-term party-to-party crossing structurally impossible there. That is a fact about a statute, recorded in the inclusion criteria as a jurisdiction-specific rule and cited here as evidence that the distinction is a live one in Canadian practice — not as an argument that it is therefore correct, and not as a proposal that any other legislature follow.

That is the whole of it. It is a view about two acts. It is not a view about any of the people in this record, and it is not a proposal that either act be forbidden.

About the act, never about the person

No record on this site says why anybody acted, and this section licenses nobody to start. The view above is about the shape of an act — where a seat went, and whether voters were asked — and never about the character, the sincerity or the calculation of a named individual. Two members can make the same move for opposite reasons; the record shows the move, and stops.

That boundary is not a courtesy, it is enforced. The published limits on machine assistance forbid approving any claim about motive, benefit, causation or disputed interpretation, and the automated verifier’s own entry in the contributor register carries the same limit in the same words. The inclusion criteria hold, as a standing rule, that the motive for a change is never part of the inclusion test. This opinion widens none of that.

What the record cannot tell you about it

The heaviest limit on this view is that the record frequently cannot say whether the member chose to leave at all.

Whether a departure was voluntary is stored on each participation, and on the large majority of them it reads unknown. That is the honest value and not a placeholder: establishing that a member walked out rather than being removed needs a source that says so, and across most of the record no such source has been cited. The live share is measured on the coverage report, which is where any figure of that kind belongs.

Expulsion is not a choice. A caucus that removes a member has acted, and that member sits as an independent because of a decision that was not theirs. Nothing in the view above applies to them: there is no act of theirs for the project to hold a view about. Because the field is unknown so often, a departure to the crossbench in this record must be read as a statement about where the member sat afterwards, and never as a statement that they decided to sit there.

The stored classification is named for that reason: it describes a seat, not a motive. The same asymmetry runs through the rest of the record. The project can often show that an election fell between a departure and a later arrival; it can much more rarely show that none did, because an election missing from an incomplete series is indistinguishable from an election that never happened. Where that is the case the record says so instead of answering.

What follows for the numbers

The consequence is arithmetic, and it is the reason the distinction was worth making at all rather than being left as a remark on a page.

The floor-crossing figure counts legs whose two sides are both parties, and different parties. It excludes departures to the crossbench, arrivals from it, moves into or out of a parliamentary group, renames, election-time relabelling, and legs on which no ruling has yet been made.
Departures to the crossbench are a figure of their own, published beside the first and never added to it.
Arrivals in a party from the crossbench count toward neither. The outbound leg of the same episode is already counted as a departure, and counting the return as well would count one person twice for one episode. It is published as a third figure so that the delayed cases stay visible rather than vanishing between the two.
Each of those figures is published together with the number of legs on which no ruling has been made, because a count that hides its unruled remainder presents itself as more settled than it is.

If you disagree with any of this

Reject the view and the data still works. The classification is stored on each participation and published through the interface, so recombining the separated figures into a single total is one addition, and rebuilding a different split is a filter. Nothing was dropped in order to hold this position, and nothing is hidden behind it: the site’s standing rule is that contested classes are recorded and flagged rather than silently excluded, and this classification is subject to that rule like every other.

The distinction is also, separately, a factual one, and it survives the opinion. Whether a seat went to another party or to the crossbench is a question about two labels, answerable without agreeing with a word of this section. The opinion explains why the project counts them apart; the classification would be worth having either way.

Text last checked against the database: 2026-09-01

Comparing our total with someone else’s

The Library of Parliament’s public figure for MPs

What they count: Persons: MPs who have changed political affiliation since 1867, split between those who moved party to party and those who were Independent before or after.

Conversion

Switch to the person unit. Count distinct persons, not events.
Restrict to the federal jurisdiction and to the House of Commons. Their figure excludes provincial legislatures; ours includes them.
Sum the right figures. Their published total is a single number covering both moves between parties and members who were Independent before or after, and roughly half of it is the second kind — so comparing it against our crossings figure alone compares one number with something under half of its counterpart. Add the crossbench departures and the arrivals from the crossbench to the crossings, then say in print that all three were summed and why, because that sum is not a figure this project publishes on its own.
Add the persons hidden inside unresolved group records. Until those records name their members, our person count sits below theirs for a reason that is arithmetic, not disagreement.

Caution: Their figure counts people who served in either chamber under a single heading. Ours holds senators but publishes none, so no senator reaches a published total. The residual after the four steps above is mostly Senate service and unresolved group membership.

A raw transition count taken from PARLINFO’s affiliation-change filter

What they count: Transitions: one row per person per change of dated party-membership role, both chambers, including Independent and non-affiliated on either side.

Conversion

Switch to the participant unit, which is the same shape as theirs: one row per person per change.
Restrict to the House of Commons, or accept that the Senate portion of their count has no published counterpart on our side. We hold Senate transitions; none of them has cleared review, so none appears in a published comparison.
Decide what to do with their same-label rows, historical relabellings and new-senator onboarding. Their raw total includes all three; our headline excludes renames, and publishes no senator at all — including the onboarding legs, which we would hold out of a crossings total even once they are published, because being appointed to a seat is not a change of allegiance.
Expand our group events into one row per participant, which is how their model would have recorded the same day.

Caution: A raw transition total from that source is higher than any journalist-grade event count, on their side and on ours. It is the right basis for machine reconciliation and the wrong basis for a sentence in a news story.

The principal public encyclopedia list of Canadian floor crossings

What they count: Entries, organised by decade and mixing jurisdictions, with a separate section for changes the list itself does not regard as crossings — chiefly mergers and pre-election party changes.

Conversion

Decide first whether to include their "other changes" section. Including it moves the comparison toward our full record; excluding it moves it toward our headline.
Compare jurisdiction by jurisdiction, never in aggregate. Their provincial coverage is uneven by era and thin before 1990, so a single total conceals which side is missing what.
Treat every difference as a lead, in both directions. An entry they hold and we do not is a candidate for research; an entry we hold and they do not is a candidate for review.

Caution: That list is a lead source, not an authority, and it is licensed share-alike. We use it to find candidates and then verify against official records; we do not copy its prose, its annotations or its row structure.

The coverage report applies the first of these recipes to today’s data.

Verification levels

What each verification level and publication status means, what a reader may rely on at each, and the standing rule that a record’s presence is not a claim that every field in it has been checked.

Two axes, not one scale

Every record carries a verification level and a publication status, and they answer different questions. The verification level says how much independent evidence stands behind the record. The publication status says how far it has travelled through editorial process. A record can be well evidenced and not yet published; it cannot be published as verified without evidence, because the database refuses to store that combination.

The refusal is a database constraint rather than a habit: a record may only carry a publication date if its status is verified and its verification level is corroborated or official and independent. No editorial decision, human or automated, can route around it.

Presence is not verification

This is the rule that matters most to anyone quoting the record. A record exists because somebody had reason to think the event happened. It does not follow that every field in it has been checked, that its date is precise, that the party names are the labels used at the time, or that the outcomes attached to it have been established.

Verification is per field and per claim, not per record. The date of an event can be verified against an official membership record while the destination party label remains a legacy value nobody has confirmed. Both states are visible on the record, and they are visible separately.

A large share of the record consists of legacy leads: entries inherited from an earlier phase of the project, retained because they point at something real, and marked unverified because nobody has yet stood behind them with a citation. They are published as leads on purpose. Hiding them would flatter the apparent quality of the record and remove the leads a researcher can act on.

Date precision is part of the fact

Every date in the record carries its own precision. A date recorded at year precision is stored with a placeholder day and month so that it can be sorted; that placeholder is not a finding and must never be printed as one. The precision is the field to read, not the date.

A great many inherited records sit at year precision, which is why a year-only entry saying that someone crossed in a given year is entirely compatible with a later, day-precise record establishing the actual date. When a day-precise record replaces a year-only one, the change is logged as a correction and the original identifier is preserved.

Claims, confidence and the limits on interpretation

Underneath the records sits a layer of individual claims: one row per assertion, each naming the record it is about, the field it concerns, its text in both languages, its status, its confidence, whether it is material, and whether it touches a living person. Citations attach to claims, not to records, which is what makes it possible to say that a date is supported while a motive is not.

Claims are typed. A statement of fact, a statement reported as having been made by someone, a claim made by a politician about themselves, an analysis by this project, a disputed interpretation and a counterfactual are six different things and are stored as six different kinds. Counterfactual claims carry the heaviest requirement in the whole schema: the database will not accept one unless a named contributor has signed it and its assumptions and counterarguments are both written out, in both languages.

Text last checked against the database: 2026-08-31

Controlled vocabulary

Verification level

Verification level
Controlled vocabulary	Definition	What you may rely on
Unverified`unverifiedaffiliation_events.verification_level`	Nobody has stood behind this record with a citation. It was inherited, or entered from a lead, and it has not been checked against a source.	Treat as a lead, not as a finding. Usable to know where to look; not usable as the basis of a published statement. Expect year-precision dates and unconfirmed party labels.
Single source`single_sourceaffiliation_events.verification_level`	One source supports the record, with a locator, and it has been read. No second source has confirmed it independently.	Usable with the source named in your own copy. Where the single source is a news report, the risk is that the report itself was the origin of a later chain of repetition.
Corroborated`corroboratedaffiliation_events.verification_level`	At least two independent sources support the record, each with a locator. Independent means they do not derive from one another.	Usable as a factual statement about the date, the parties and the person. Motive, benefit and causation are never covered by this level, whatever the sources say about them.
Official and independent`official_and_independentaffiliation_events.verification_level`	An official record — a chamber’s membership record, a members’ profile, a journal or an election authority — and at least one independent account agree on the literal facts.	The highest level the project awards. Usable as a factual statement without attribution to this site, though a citation is appreciated. Still says nothing about why the person moved.

Publication status

Publication status
Controlled vocabulary	Definition	What you may rely on
Legacy lead`legacy_leadaffiliation_events.publication_status`	Inherited from an earlier phase of the project. Published because it points at something real and someone may be able to resolve it; not published as a finding.	The label is a warning, not a formality. Legacy leads are the part of the record most likely to carry a wrong year, a collapsed episode, or a party label that did not exist under that name at the time.
Pending review`pending_reviewaffiliation_events.publication_status`	Entered or amended and waiting for editorial review. Sources may already be attached; the review has not concluded.	Read the attached sources yourself. What is pending is our judgement, not necessarily the evidence.
Verified`verifiedaffiliation_events.publication_status`	Reviewed and published as a finding. Only records at corroborated or official-and-independent verification may reach this status with a publication date.	Citable. Check the date precision before writing a day into a sentence, and check the verification level to know which of the two evidentiary standards was met.
Disputed`disputedaffiliation_events.publication_status`	Sources of comparable standing disagree, or a party to the events has contested the record and the contest has substance. The disagreement is published rather than resolved by preference.	Do not pick a side from this record. Read both sources; if you publish, publish the disagreement.
Withdrawn`withdrawnaffiliation_events.publication_status`	The record does not meet the inclusion criteria, or was wrong, and has been retired. It is kept at its original identifier, retitled so that the withdrawal is legible, and linked to whatever replaced it.	Not part of any count. Present so that a citation made before the correction resolves to an explanation instead of to nothing.

Date precision

Date precision
Controlled vocabulary	Definition	What you may rely on
Day`dayaffiliation_events.date_precision`	The exact calendar day is established by a source.	The only precision from which a day may be printed.
Month`monthaffiliation_events.date_precision`	The month and year are established; the day is not.	Print the month. The stored day is a sorting placeholder.
Year`yearaffiliation_events.date_precision`	Only the year is established. The stored month and day are placeholders that exist so records can be ordered.	Print the year alone. This is the most common precision among inherited records and the most common source of a wrong date in downstream work.
Circa`circaaffiliation_events.date_precision`	The date is approximate and the sources do not pin it down.	Print with the approximation intact. Do not silently harden it into a year.
Range`rangeaffiliation_events.date_precision`	The change occurred somewhere within a stated interval, with both ends recorded.	Print both ends. A midpoint is an invention.
Unknown`unknownaffiliation_events.date_precision`	The date is not established at any precision.	The record cannot support a date in any form.

Claim status

Claim status
Controlled vocabulary	Definition	What you may rely on
Pending review`pending_reviewclaims.assertion_status`	The claim is recorded and has not yet been assessed against its citations.	Not a finding. Most claims inherited from the legacy record sit here.
Supported`supportedclaims.assertion_status`	Cited sources support the claim as written, at the locators given.	Citable, with the sources named. Check the confidence value alongside it.
Partially supported`partially_supportedclaims.assertion_status`	Part of the claim is supported and part is not. The unsupported part is named in the claim’s own text.	Quote only the supported part, and read the claim text to find out which part that is.
Disputed`disputedclaims.assertion_status`	Sources of comparable standing contradict each other on this specific claim.	Publish the disagreement rather than a resolution.
Unsupported`unsupportedclaims.assertion_status`	The claim was assessed and the cited sources do not support it. It is kept, marked, so that the assessment itself is on the record.	Do not cite. Its presence is a finding about the claim, not evidence for it.
Unknown`unknownclaims.assertion_status`	The evidence available does not settle the claim either way.	An honest null. Not the same as unsupported, and not the same as pending.
Withdrawn`withdrawnclaims.assertion_status`	The claim has been retracted, with the reason logged in the corrections ledger.	Retained so that anyone who relied on it can find out why it went.

Sourcing standard

Source tiers, the requirement that a citation carry a locator rather than an address alone, how sources are archived and monitored, the prohibition on machine-generated facts, and a full account of the fabricated content this project once carried and removed.

A URL is not a citation

A citation in this record points at a specific place inside a specific source. It carries a locator — a page, a paragraph, a section, a record number, or a timestamp — alongside the source itself. An address that merely points at a document, leaving the reader to find the relevant passage, is not a citation and does not raise a record above unverified.

Citations attach to claims rather than to records. A single event can hold a citation supporting its date at a locator in an official membership record, a second supporting the destination party at a locator in a news report, and no citation at all for the member’s stated reasons — and the record shows all three states side by side rather than averaging them into an overall impression of reliability.

A citation can also record that a source contradicts the claim it is attached to. Contradicting sources are stored, not discarded, because the fact that a reputable source disagrees is itself a finding a reporter needs.

Archiving and link rot

A source that has disappeared cannot be checked, and a historical record whose citations decay becomes unverifiable at exactly the speed of the web. Every source in the record therefore has room for an archive address, the date that archive copy was captured, the date the source was consulted, and a checksum of what was read.

Between captures, links are monitored automatically. Each check records the date, the response, the final address after any redirection, and a status: reachable, redirected, broken, blocked, changed, or unknown. A source whose content has changed since it was cited is as serious as one that has vanished, and is flagged separately.

What we do with the sources themselves

Every source carries a rights status, and every source in the record currently sits at link-only. We link to sources; we do not republish their text. Where an excerpt is needed to show that a source says what we say it says, it is stored as a short quoted passage attached to the citation, not as a copy of the document.

Parliamentary material carries one hard condition from its own terms, and we observe it: reproductions of proceedings must be accurate and must not be presented as official. Nothing on this site is an official publication of the Senate, the House of Commons, a provincial legislature, or the Library of Parliament, and nothing here is endorsed by them.

No machine-generated facts

No fact, date, name, quotation, vote count, margin or narrative in this record may originate from a language model. Machine assistance is permitted for finding candidates, for matching entities, for reading scanned text and for drafting translations, and every use of it is recorded on the row it touched. It is never permitted as a source.

The rule has a corollary that governs everything else in this document: an empty field beats a plausible one. Where a value has not been established, the record carries no value and says so. Filling a gap with something that reads correctly is the failure this project has already made once, and it is the failure that costs a reference work its standing.

Provenance and past errors

An earlier phase of this project published fabricated content. It was written by a language model directly into the database, without citations, and it stayed up for months. The account below is complete as far as we have been able to reconstruct it, and the scripts that did it remain in the repository where anyone can read them.

What was invented. Election results were fabricated outright, and the same numbers were reused across unrelated people: one five-digit vote total appears for eight different members across seven elections spanning fifty-three years, from 1972 to 2025. Forty-four direct quotations were entered with no source address of any kind, attributed to press conferences and parliamentary debates that were never identified. One hundred and sixteen narrative timeline entries were entered the same way, several carrying placeholder dates of 1 January. Long contextual essays and biographies were produced by the same process.

Errors inside the invention. Some of the fabricated prose was not merely unsourced but internally impossible: one essay described a confidence vote as carried 153 to 152 with the Speaker breaking the tie, which cannot occur, since a Speaker breaks a tie only when the vote is tied. Two scripts disagreed about which Parliament a 2026 crossing belonged to. A name-matching routine attached an American congressman’s biography and portrait to a Manitoba member, and a hockey executive’s portrait to a British Columbia legislator. A generated passage asserted that one member crossed following another when the order was in fact reversed.

What was removed. A purge deleted every quotation, every narrative timeline entry and every source row in the database, nulled every vote count, vote share and margin, and cleared every generated biography and contextual essay. The script that did it states in its own header that the numbers were machine-generated and not real data. A later script then reintroduced some of the same material, including the impossible vote tally, which was removed again during the rebuild of the data model.

What replaced it. Narrative prose asserting facts no longer exists as a field. Facts are stored as individual claims, each bound to citations with locators, each carrying its own status and confidence. Election results are not restored until an election authority’s record supports them, which means the record currently shows fewer numbers than it once did. That is the correct direction of travel.

What is still open. The portrait layer is not yet clean: images remain whose only rights basis was another site’s internal non-free-content policy, which does not transfer here. Archive capture has not run. A majority of records remain legacy leads at year precision. None of this is concealed, and each is stated where it bears on what a reader can rely on.

Why this is published. Because the alternative is worse, and because a reference work that has never had to purge anything is usually a reference work that has never audited itself. The controls below exist because of specific failures listed above, and each one names the failure it is there to prevent.

Text last checked against the database: 2026-08-31

Source tiers

Source tiers
Controlled vocabulary	Definition	What you may rely on
Primary official`primary_officialevidence_sources.reliability_tier`	The institution’s own record of its own business: a chamber’s membership record or journal, a members’ profile maintained by a legislature, an election authority’s returns, a gazette.	The strongest basis for a date, an affiliation, an office or a result. Silent on motive, and often later than the event it records.
Primary, non-official`primary_otherevidence_sources.reliability_tier`	A first-hand document that is not an institutional record: a party’s own announcement, a member’s letter or statement, a contemporaneous photograph of a document.	Authoritative about what its author said, not about whether what they said was true.
Secondary, reliable`secondary_reliableevidence_sources.reliability_tier`	Reporting by an outlet with an editorial process and a corrections practice, published close in time to the events.	The usual corroborating source for a date, and often the only source for what was said at the time. Two outlets repeating one wire report are one source, not two.
Scholarly`scholarlyevidence_sources.reliability_tier`	Peer-reviewed research, a scholarly dataset, or an academic monograph.	Strong for classification and for context; check the coverage window before relying on it for a specific case, since most Canadian party-switching scholarship stops well before the present.
Tertiary`tertiaryevidence_sources.reliability_tier`	Encyclopedias, aggregated lists and reference compilations that summarise other sources.	A lead source. Never sufficient on its own to raise a record above unverified, and never the basis of a published statement of fact.
Unassessed`unassessedevidence_sources.reliability_tier`	The source is attached but nobody has judged what kind of source it is.	An honest null rather than a default. Read it yourself before relying on it.

Locator types

Locator types
Controlled vocabulary	Definition	What you may rely on
Page`pageclaim_citations.locator_type`	A page number in a paginated document.	Used for journals, reports and printed debates.
Paragraph`paragraphclaim_citations.locator_type`	A numbered paragraph, where the document numbers them.	Used for decisions, rulings and structured reports.
Section`sectionclaim_citations.locator_type`	A named heading or subdivision within a document or page.	Used for web documents that have no page numbers.
Timestamp`timestampclaim_citations.locator_type`	A time offset in a recording or broadcast.	Required for any claim resting on video or audio.
Record`recordclaim_citations.locator_type`	An identifier inside a structured system: a member profile, a role period, a return, a file number.	The strongest locator, because it survives redesigns of the source’s website.
Other`otherclaim_citations.locator_type`	A locator that fits none of the above and is written out in full.	Acceptable. A missing locator is not.

The controls against fabrication

Facts are stored as individual claims bound to citations. A claim with no citation cannot reach supported status.

Enforced by: The claims and citations tables, and editorial review.
Prevents: Free narrative prose asserting facts that nothing supports, which is how the original fabrication entered the database.

A record may only carry a publication date if its status is verified and its verification level is corroborated or official-and-independent.

Enforced by: A database constraint. No process, human or automated, can route around it.
Prevents: Unverified material being published with the appearance of having been checked.

Vote counts, shares and margins come from an election authority’s record or they are absent. Empty beats wrong.

Enforced by: Editorial rule, and the removal of every inherited number that lacked such a record.
Prevents: Invented vote totals, including the recycling of one figure across unrelated politicians.

A quotation is stored as a reported statement, a claim kind distinct from fact, and requires a locator inside a source that carries the words.

Enforced by: The claim-kind vocabulary and the locator requirement.
Prevents: The forty-four unsourced quotations attributed to unidentified press conferences and debates.

Motive, benefit and causation are never stored as facts. An outcome carries an explicit causation status whose default is that causation has not been assessed, and there is a distinct value for an official finding that no special inducement occurred.

Enforced by: The outcomes vocabulary, and the rule that no automated actor may set it.
Prevents: A dated appointment beside a dated crossing being published as a reward, which the earlier scoring layer effectively did.

Automated actors are contributors with published limits. The automated verifier’s own biography states that it validates literal dates, affiliations, offices and results against explicit locators in independent official systems, and that it is not authorised to approve motive, benefit, causation, disputed interpretation or sensitive biographical claims.

Enforced by: The contributor record, the review ledger, and process. This one is a published commitment rather than a database constraint, and is named as such.
Prevents: A machine signing off on the interpretive claims that caused the original problem.

A counterfactual claim cannot be stored at all unless a named contributor has signed it and its assumptions and counterarguments are written out in both languages.

Enforced by: A database constraint.
Prevents: Speculation about what would have happened being stored with the same weight as a dated fact.

Every change writes a revision naming the fields it touched; every material correction is published in the corrections ledger and bound to that revision.

Enforced by: The revisions table, and a schema requirement that a correction reference a revision.
Prevents: A repeat of the episode in which fabricated material was removed, quietly reintroduced by a later script, and removed again, with no public record of any of it.

Loading…