Field OP — Labelling Terms

Version: 1.0-draft Published at: /terms/labelling and /terms/labelling/1.0-draft, and at Admin → Labelling in the console Effective: not in force. This version is engineering's, no lawyer has reviewed it, and the page carrying it says so before the text.

These are the terms an administrator agrees to before Field OP will label anybody in your organization's records from their family name.

One text, in two places

This document is section 9 of the Terms of Service, whole, and with section 9's numbering kept. It is not a summary of section 9 and it is not a plainer telling of it. It is the same words, so that what your organization agreed to when it signed up and what an administrator agrees to on the labelling screen cannot drift into two different promises. Section 16.4 of the Terms of Service says the two are one text; pnpm check:legal is what makes that true rather than intended, by comparing the two files byte for byte and failing the build on any difference between them. Changing a word here is therefore a new version of both documents, which is the honest consequence of their being one text.

The numbering below is section 9's numbering, so that a citation means the same thing in both places.

One thing is deliberately not repeated here. Section 5.5 of the Terms of Service sets out the pooled calibration counts — what crosses the boundary between organizations when the model's guess and a person's own answer are compared, what a count can and cannot say, and the switch your administrators hold. It is linked rather than restated, because another copy of a text is another thing to keep in step, and the switch itself sits on the screen this document is shown beside.

Section 9.1 below carries a note saying that this page and the Inferred Trait Schedule do not exist. They exist, and have since 18 September 2026. That note is frozen inside terms/1.0-draft.md, which cannot be edited without invalidating every acceptance recorded against it, and it is reproduced here unaltered — because reproducing it altered is precisely what reproducing it is meant to rule out. The correction is recorded against the next version in the publication checklist.

Section 9 of the Terms of Service

9.1 What the feature is

Field OP includes a model that reads a person's family name and records what it guesses about them. It is a guess from a name and nothing else — not from anything the person told you, and not from anything they did.

The characteristics the model guesses are a person's likely language, ethno-cultural community and religion, and the further characteristics set out in the Inferred Trait Schedule, which forms part of these terms.

[NOT YET BUILT — 17 Sep 2026] The Schedule and the Labelling Terms are addressed below as /terms/trait-schedule and /terms/labelling. Neither page exists: the console serves /terms, /privacy, /acceptable-use, /data-processing and /subprocessors, and nothing else. Both documents are real and are part of this set; they are not yet served at an address. We may add to or change the Schedule on notice to your administrators; a change to the categories named in this section is a change to these terms and is handled under section 16.3.

9.2 On by default, and you can switch it off

Labelling is on for a new organization by default outside Québec, as a matter of Field OP's platform policy. An organization that gives a Québec address when it signs up starts with labelling off and no characteristic enabled at all.

Your organization's settings row records the founding administrator as the person it was turned on by, and records that it was turned on at sign-up under the version of these terms that was accepted there — not under the Labelling Terms, which nobody had been shown at that point. The screen at Admin → Labelling says exactly that, and invites an administrator to turn the feature off and on again if they want their agreement recorded against the Labelling Terms themselves.

A tenant administrator may switch it off at any time, and labels stop being produced immediately. Turning it off and on again records that somebody agreed to the terms then in force. Labels already written stay on record, so that the decision can be audited and anything acted on can be found; they can be deleted, but only as a deliberate act.

Field OP does not represent that this feature is lawful in your jurisdiction, and you should take your own advice before relying on it.

9.3 What a label is never allowed to be

• A guess is never recorded as a fact. A label is stored apart from the person's record, marked as inferred, with the model, the model version, the evidence, the explanation and a date it stops being current. It is never written into a name field, a language field or anything else your organization holds as a fact.
• A trait a person states about themselves outranks the model's guess, always. If somebody tells your volunteer their language, that is what your organization holds; the guess does not overrule it and does not come back.
• No message is ever selected from an inferred trait. A script variant may select only on fields the mission bundle already carries plus the language a person stated; an inferred trait structurally cannot be one of them.
• No one is included in or excluded from contact on a score. The rule language that would let an organization build a target list from model output is switched off everywhere and matches nobody.
• A human can confirm, correct, reject or switch off any label, for any individual, and a label a human rejected is not produced again by the same version of the model. Low-confidence guesses are stored and never shown.
• Who can never see a label: no volunteer, no team captain, no candidate — not on a screen, not on a phone, not in an export, not on a printed walk sheet. Reading a label requires the Restricted Data grant, which is dual-controlled, expiring and separate from any role. This is enforced in the database, so it holds for software written later, and an export refuses an inferred class before any grant is considered.

9.4 The marker on every label

Every label carries a marker recording the platform policy, the model and the model version under which it was computed, so that any label can be traced to the rules and the wording in force when it was written.

9.5 How wrong it is

It is wrong often, and unevenly. On the synthetic test population the model is right about roughly 88% of names from one community and 62% from another — twenty-six points apart, where a single average would report seventy-eight percent and conceal both ends. The accuracy report is ordered worst-community-first for that reason, and your administrators can see it before and after switching the feature on.

Religion is the weakest. A name can suggest a community. It cannot tell you what somebody believes, or whether they believe anything. Field OP produces a religion label by applying a published assumption about a community to a person, and usually declines to name one answer and shows a spread instead. Treat it as the least reliable thing the platform will ever show you.

A name is weakest evidence precisely for the people most often misunderstood — adoptees, converts, people who married across a community, and families longest in the country.

On a real campaign the accuracy report cannot break down by community at all, because nobody knows anybody's actual community. Only a person's own correction can measure it, one at a time, after the fact.

9.6 Nothing is sent anywhere

The model runs inside Field OP on Canadian infrastructure. No name, and nothing else about any person, is sent to a third-party name-classification service, to a demographic-inference API, or to a general-purpose AI provider, for this feature or any part of it.

9.7 Human review, and your responsibility for what you decide

This applies to every AI feature in the Service, not only labelling.

• AI output is a suggestion, and a human confirms it. Import column mapping is proposed by a model and accepted, edited or rejected by a person, and the outcome is recorded. Anything a model produces about a person is an estimate.
• You are responsible for your targeting and messaging decisions, including decisions you make using model output. If you choose who to contact, what language to offer, or what to say on the basis of something Field OP inferred, that is your decision and your organization's responsibility, and you are responsible for it being lawful and non-discriminatory.
• You will use labels only for lawful, non-discriminatory purposes connected to your stated work — offering material in a language somebody is likely to read is the example the product gives. You will not use them to exclude anyone from contact, to target anyone on the basis of religion, or for any purpose that would be unlawful if the person had told you the same thing directly.
• If a person asks what your organization believes about them and why, that answer is yours to give. The platform can produce it — the label, its confidence, the model version, and the fact that it came from their name — and there is no self-service route for a member of the public. They come to you.

What agreeing here records

Turning labelling on from Admin → Labelling records labelling:1.0-draft against your organization's inference settings, with the administrator who agreed and the moment they agreed. Every label the platform then writes carries that authority. Switching the feature off and on again records agreement to whichever version is in force at that moment.

An organization provisioned outside Québec starts with labelling already on, and its settings record sign-up: followed by a version of the Terms of Service instead of a version of this document — because section 9 is what the founding administrator accepted at sign-up, and nobody had been shown this screen at that point. The screen says exactly that on its face, and turning the feature off and on again is how an agreement comes to be recorded against this document.

An earlier and shorter set of these terms was shown in the console between 11 and 18 September 2026 and recorded as name-labelling-2026-09-v1. It is kept, at /terms/labelling/0-draft, so that a setting recorded against it still resolves to the words somebody read.

By enabling or continuing to use this feature, your organization accepts these terms. Your acceptance is recorded against this version.