Boundary Sets: A Field-by-Field Setup Guide

Most consent problems are not disagreements about values. They are gaps between what someone meant and what a system could act on. A likeness boundary set closes that gap by turning your position into fields a machine reads on every request — before anything is generated, not after someone complains.

This is the field-by-field walkthrough. It covers each part of the set, what a good default looks like, and the specific settings creators most often get wrong. If you have not yet decided whether to license at all, start with how to license your likeness and come back.

TL;DR

• Five parts: permitted contexts, forbidden depictions, realism ceiling, conversation mode, relationship framing.
• Contexts are allow-listed — anything you do not tick is refused.
• Four prohibitions are refused platform-wide and evaluated before your rules.
• Sets are versioned append-only, so old outputs stay traceable to the version that authorised them.
• Refusals name the rule they broke. If yours do not, the enforcement is probably cosmetic.

On this page

• Where the settings live
• Field 1: permitted contexts
• Field 2: forbidden depictions
• Field 3: realism ceiling
• Field 4: conversation mode
• Field 5: relationship framing and session caps
• How the evaluation actually runs
• A worked example
• Pitfalls worth avoiding

Where the settings live

Sign in, open the agency console, pick a creator profile, and choose Boundaries. Every change creates a new version rather than editing the old one — you will see the version number tick up when you publish. The same data is readable publicly at /api/v1/creators/{slug}/boundaries, which is how a buyer or an agent checks your rules before spending anything.

That public readability is the point. A rule nobody can inspect before requesting is a rule that only exists at the moment of refusal, which wastes everyone's time.

Field 1: permitted contexts

Contexts describe where your likeness may appear: portrait, fashion editorial, fitness, gaming avatar, brand endorsement, music performance, travel and lifestyle, seasonal, cosplay, and so on. Full definitions sit on the boundary reference.

The critical property: this list is an allow-list. An empty list means nothing is licensable. A request for a context you did not tick is refused with context_not_permitted, naming the context it asked for.

Start narrow. Portrait alone is a perfectly sensible v1. Every context you add widens the surface, and widening later is easy while narrowing later means explaining to existing subscribers why something stopped working.

Which contexts should I enable first?

If you are a model or style creator, portrait plus editorial and fashion. If you stream, portrait plus gaming and avatars. If you have sponsors, add brand endorsement and fill in the excluded-brands list in the same sitting — the two belong together and splitting them across two days is how the second half gets forgotten.

Field 2: forbidden depictions

Depictions describe what is shown. This list is a deny-list layered on top of the contexts you permitted, and it wins any conflict.

Table: the forbidden-depiction categories most often adjusted during first setup, and the reasoning creators give. The four platform-wide prohibitions are omitted because they are not configurable.

Two of these carry statutory weight rather than just preference. Sexual content is where the Online Safety Act 2023 age-assurance duties and the TAKE IT DOWN Act removal duty both land, so forbidding it narrows the platform's exposure as well as protecting you.

Note the fourth column that does not exist: there is no "ask me first" state. A depiction is forbidden or it is not. Approval queues sound reassuring and in practice become a backlog somebody clears at speed on a Friday.

Field 3: realism ceiling

Three values — illustrative, stylised, photoreal — and the request must sit at or below yours. Ask for photoreal against a stylised ceiling and you get realism_exceeds_ceiling, with both values quoted so the refusal is self-explanatory.

This is the single most effective control in the set, and it is the one most often left at maximum. Harm concentrates in output that could pass for a photograph. If you are hesitating, set it to stylised. Raising it later takes one edit.

Field 4: conversation mode

Separate from imagery: whether your likeness may be used for conversational features at all. Off by default. A request with mode=conversation against a creator who has not licensed it is refused with conversation_not_licensed.

Keep it off unless you have thought specifically about it. Conversational use is a different product with different attachment dynamics, and licensing imagery does not imply licensing a voice that claims to be you.

Field 5: relationship framing and session caps

This is the part with no equivalent on unlicensed platforms, and the part creators ask about most once they notice it.

You can set session caps and variety nudges: after a set number of consecutive generations of you, the platform nudges the requester — a soft prompt by default, a harder throttle if you turn it up. It runs on the same server-side path as the content rules, so it is an actual mechanism rather than a wellbeing paragraph in the terms.

It ships on with a soft nudge. That default is a compromise and we would rather say so than pretend otherwise: a hard block would protect attachment dynamics better and would also reduce revenue, including yours. Making it a creator-set lever puts that trade in the hands of the person whose face it is. We looked at the evidence and the incentives in the engagement data study.

How the evaluation actually runs

Order matters more than most people expect, so here it is plainly:

1. Platform prohibitions — the four non-waivable categories. Checked first.
2. Creator rules — your contexts, depictions, ceiling, conversation flag.
3. Relationship framing — session caps and nudges.
4. Age assurance — the requesting account's verification state.
5. Credits — is this request paid for?

Prohibitions run before the age gate, so a prohibited request from a fully verified adult account is still refused. Credits run last, so a request that breaches a rule is never billed. That ordering is the difference between a consent engine and a payment funnel with filters attached.

Text is normalised before matching — accents, homoglyphs, full-width characters, spaced-out letters and digit substitutions are folded first, and the refusal quotes the decoded term. In plain terms, spelling a forbidden word oddly does not get it through.

A worked example

An agency onboards a fitness creator. Her position, in her words: workouts and sportswear fine, nothing sexual, nothing photoreal, no supplements.

That becomes:

• Contexts: portrait, fitness, fashion editorial, travel and lifestyle.
• Forbidden: nudity, sexual content, medical/health claims, financial endorsement, intimate partner framing.
• Ceiling: stylised.
• Conversation: off.
• Session cap: nudge after 12 consecutive generations.

Published as v1. Three weeks later a supplement brand signs her, so the agency adds competitor brands to the excluded list and publishes v2. Everything generated under v1 still records v1 — the audit trail does not blur, and if a dispute lands in month nine, the question "what were the rules on the day?" has an answer rather than a debate.

Pitfalls worth avoiding

Enabling every context on day one. It feels generous. It mostly generates refusals you did not anticipate and support tickets you did not want.

Leaving the realism ceiling at photoreal by default. Covered above, and worth repeating because it is the most common single mistake.

Filling in excluded brands "later". Later is after the campaign.

Assuming forbidden depictions imply a low ceiling. They are independent. Forbidding nudity does not stop photoreal portraits.

Copying a house template across a whole roster. Convenient, and it quietly means one person's judgement is standing in for ten people's consent. If you manage creators, at minimum walk each one through their own set once.

Not testing a refusal. Submit something you know breaks a rule and read what comes back. If the message does not name the rule, you cannot audit the system, and neither can a regulator asking you to.

---

Next: see what the enforcement produces in practice on the provenance explainer, or, if you are integrating rather than configuring, the licensed-likeness API tutorial. Agencies managing several profiles should start at the agency overview.