Labels and what they mean

Field OP can guess some things about a person from their record — most often the language they are likely to speak. A guess is never a fact, it is never mixed in with the facts, and nobody at a door ever sees one. This page says how it is marked, who may read it, and how to correct it.

It is on for an organization that signs up, and an administrator can turn it off

Labelling is on for an organization that signs up outside Québec, and off for one in Québec or one Gestalt sets up for a campaign directly. Admin → Name labelling says which, and whether it was switched on by the platform at sign-up or by an administrator who read the labelling terms and agreed to them on behalf of the organization — the version they agreed to is recorded against their name.

A guess about somebody's faith starts off for every new organization, however it was set up, and even where the rest is on — turning labelling on does not turn faith on with it. The step from a community to what somebody believes rests on a table nobody could yet stand behind, so faith is guessed only if somebody in the organization turns it on by name. An organization that turned it on before this change keeps what it chose.

That screen leads with a warning, and this page repeats it rather than softening it: this feature is not fit for a real campaign yet. A legal opinion has been asked for and has not arrived.

On a phone the same screen is in the Field OP app, under More → Admin → Name labelling, and the console's screen works at a phone's width too. It shows the same switch and the same warning first; the terms open over the whole screen, every word of them, with the documents they name — the Inferred Trait Schedule and the Terms of Service — a tap away, and agreeing records the version you read exactly as it does on a computer. The platform records an agreement only to the terms in force, so an app built before the terms changed says to update it rather than recording the old ones. Each control is there only for the account that may use it — turning it off, switching off a kind and stopping the sharing of counts for the tenant administrator, turning it on, adding a kind and moving the floor for an administrator who also holds the Intelligence grant, running it for somebody holding Restricted Data. Nothing on these screens names a person or shows anybody's guess: they are the switch, the counts a run reports, and the accuracy report.

A tenant administrator can always see that screen and turn labelling off, switch off a kind, or stop sharing counts — with their role alone, because turning something off widens nothing. Turning it on, adding a kind, or moving the confidence floor (a guess the model is less sure of than the floor is kept but shown to nobody) also needs the Intelligence grant — and the grant counts only for a member of staff: given to a volunteer, a captain, a candidate or an observer, it does nothing at all. Somebody holding the Restricted Data grant can run the model on demand for a campaign they work on; running it for the whole organization is the tenant administrator's alone.

A run reaches everybody in the organization, whether it starts after an import or from the button. It works through a couple of thousand people at a time and saves as it goes, so a large list takes longer rather than stopping short. From the button you can stop it and, while the page is still open, continue from where it stopped; if somebody else's run is labelling the same campaign at that moment, the page says so and you can continue once it has moved on. Running it again over people nothing has changed about keeps their guesses rather than writing them again.

What a guess looks like on a record

On a person's record, guesses live in their own panel, What the software guesses, always below the facts. Its first sentence is the marker:

Guessed from this person's family name by a model. Not facts, and not what they told us.

Every row carries the word guessed with the date, an expiry date, the model's name and version, and how confident it was. Where the model did not settle on one answer, the whole spread is shown as a list of percentages rather than collapsed to its leader — because "sixty per cent this, thirty per cent that" is a different thing to know than "this".

A guess that no longer has any support is not quietly kept: it expires. Past its expiry date a guess is shown to nobody, and it is marked out of date. Every guess names the lists it was made from, so that when a list is destroyed every guess made from it can go out of date the same moment. The console has no way to destroy a whole list yet; when it does, that is what will happen.

A family name that reached Field OP only on the voters' list from the elections agency is never read for a guess about somebody's language, community or faith.

Who can read one

A volunteer, a captain or a candidate cannot, ever, by any path — not on a record, not in the bundle their phone downloads, not in a search, not in an export. They are refused in words rather than shown an empty panel. This is a rule of the platform, not a setting a campaign can change.
An observer cannot either, whatever grant they hold: an observer reads a campaign's ordinary records and no restricted ones.
Everybody else needs the Restricted Data grant, and reads only the guesses of the campaigns that grant was approved for. Guesses made for the whole organization are read only by the tenant administrator, holding a grant that names no campaign.
Every read is recorded — who looked at which person's guesses, and when, but never what the guesses were.
On a phone too, for the same accounts. The console on a phone and the Field OP app show the same panel to an account that may read guesses, read afresh each time the record is opened and never kept on the phone; the app stops drawing it while it is not in front. A volunteer, a captain, a candidate or an observer never sees it, on any screen. Restricted details — a full date of birth, sex or gender — are still not shown on a phone.
Nothing inferred is ever in an export file.

If the campaign has not turned labelling on, the panel says nothing is guessed about this person.

Correcting one

Four buttons sit on every row:

This is right — confirm it.
Correct it — say what it should be. The form is explicit: this is kept as a person's answer, not as another guess. You can add how you know, and that is kept too.
This is wrong — reject it.
Hide it — keep it out of the way without saying anything about whether it was right.

A correction supersedes the guess: the model's own row is rejected and kept beside it for the record, the corrected row is marked as corrected by a person, and the same version of the model will not guess that person again. Nothing is rewritten and nothing is deleted — the trail of what was guessed, what was corrected and by whom stays whole.

What the campaign learns from corrections

When a guess is corrected, the pair — what the model said and what turned out to be true — can be counted, so the model can be measured. By default those counts are pooled across the organizations that share them. The counts hold no person and no campaign name, and another organization's counts reach you only inside a group that enough other sharing organizations contributed to, so nothing about one campaign can be read out by subtracting your own. Your own numbers are never withheld from you. An administrator can switch sharing off on the labelling screen.

The accuracy report

Beside the queue is an accuracy report, per label and per group, ordered worst first. It deliberately refuses to print an average: an average over communities hides exactly the number somebody needs to see. On a phone each community is a card, still worst first, and the report there names no overall figure at all — only how far apart the best-served and worst-served communities are. It is read behind the same Restricted Data grant as the guesses themselves, and every read of it is recorded.

Next: Privacy, do-not-contact and rights requests · People and imports · Getting help