What the law asks for, and what answers it
Three regimes converged on the same requirement in the space of two years: prove consent, and be able to act on its withdrawal quickly. Every mechanism below exists to answer a specific one of them.
Answered by the takedown intake, which is open to anyone without an account and stamps a due time 48 hours from the moment a report lands rather than from when somebody opens the queue. The agency console sorts by that due time and marks overdue items.
The reporting route →Answered by a document-based check through Stripe Identity, with the age flag set from the verified date of birth rather than from the session merely completing. A session that verifies a 16-year-old is a successfully verified minor; treating that as a pass would defeat the gate. With no verification provider configured, every generation is refused.
Answered by the boundary set plus a signed manifest per output. Produced at the moment of authorisation, not reconstructed when a query arrives — and verifiable by the party asking, without taking our word for it.
Verify a manifest →Why enforcement sits before the model
Post-hoc moderation has a structural problem in this category: by the time a reviewer sees the output, it exists. Somebody has it. The remedy is takedown, which is a race against distribution rather than a control.
Moving the check in front of the model changes what the platform is asserting. Instead of “we remove what breaches the rules when we find it”, the claim is “output that breaches the rules was not produced” — and the refusal, with its rule and matched term, is recorded as evidence that the rule was applied.
This is not a claim that nothing can slip through. Lexical matching can miss a phrasing. What does not depend on wording is the rest of the gate: the context allowlist, the realism ceiling, the verification state of the requester, the revocation status of the boundary set. Those are structural, and they are why the refusal path is not the only control.
The four categories no creator can license away
Most of the boundary vocabulary is the creator’s to decide. These four are not, because in each case the creator’s permission is not the only consent that matters, or the harm lands on somebody who was never asked: minors-context, non-consensual-framing, deceptive-news-context, criminal-context. A boundary set that tries to permit one is rejected when it is saved, not silently trimmed.
Questions
- What does the TAKE IT DOWN Act require of a platform?
- In plain terms: a way for a victim to notify the platform, and removal of the reported material — and known copies — within 48 hours of a valid request. The hard part is not the 48 hours; it is knowing what to remove. That is answerable only if every output can be traced to a record, which is what the provenance manifest is for.
- What does the UK Online Safety Act require?
- Among other duties, age assurance before access to content that is harmful to children, and an accessible reporting route. LikeWard gates every generation behind a verified adult account rather than a self-declared age, and the reporting route needs no account at all.
- What are card networks asking for?
- Documented consent from every identifiable individual depicted, producible on request. A signed boundary set plus a manifest naming the version it authorised is that document — produced automatically per output rather than assembled after a query lands.
- Does any of this remove the need for human review?
- No, and nothing here claims it does. What it removes is the position where a platform cannot answer what it authorised, for whom, and under which rule. Review is better when the record exists; it is not replaced by it.
This page describes what the product does and why. It is not legal advice, and the summaries of each regime are deliberately short — read the instruments themselves, or take advice, before relying on any of it for a compliance position.