TUTELIQ AB · PUBLISHED FRAMEWORK · TUT-CSDT-001
A disclosure schema for automated child safety detection systems.
DOCUMENT
TUT-CSDT-001 v1.0
STATUS
Working draft for public comment
PUBLISHED
August 2026
LICENCE
CC BY 4.0
Approximately ten minutes. Scored in your browser; no data is transmitted.
§1 · RATIONALE
You cannot regulate, procure, or improve what nobody can describe
Every platform that serves children says it uses AI to keep them safe. Almost none will say what it detects, at what precision, in which languages, on what evidence, or what happens to a child when it gets the answer wrong.
This is not usually dishonesty. The measurements mostly do not exist. "Accuracy" gets quoted where precision and recall are what matter. Detection is described as covering thirty languages when performance has been measured in two. Grooming, a relational process that unfolds over weeks, is detected message by message, which is not detecting grooming.
Regulators are converging on this problem from several directions at once. Ofcom's technology notices regime requires accredited technology to meet minimum standards of accuracy that the UK Government has not yet set. The EU AI Act brings detection systems into scope from December 2027. The interim measure permitting voluntary CSAM detection in the EU expires in April 2028.
§2 · SCOPE AND EXCLUSIONS
What this document is, and is not
THIS IS
A disclosure schema. Six pillars, roughly eighty questions, stable clause identifiers, a machine-readable JSON Schema, and a validator.
THIS IS NOT
A compliance standard, a certification, a safety rating, or a substitute for legal advice. It has no regulatory standing, it is published by a commercial vendor, and it is not yet suitable for adoption as a standard.
IT DOES NOT COVER
Age assurance, risk assessment, governance and accountability, user reporting, recommender safeguards, default settings for child accounts, or terms of service enforcement. This covers detection capability only.
§3 Structure: the six pillars
Detection coverage
Which of thirteen harm categories does your system address, by what method, and where are the gaps?
Detection quality
Precision and recall at a declared operating point, with confidence intervals and prevalence-adjusted figures. "Accuracy" is not accepted; it is trivially satisfiable for rare events.
Modality coverage
Text, image, video, audio, documents, synthetic media, cross-modal. Disclosed as a harm by modality matrix, not a list.
Language and culture
A language is not supported because the model can process it. It is supported when performance has been measured and disclosed.
Data protection and lawful basis
Article 6 basis, Article 9 condition, DPIA, ePrivacy, the Children's Code, Article 22, and the lawful handling of illegal material.
Human oversight and outcomes
What happens after a detection event, what happens to the user, and what the appeal reversal rate says about how often you are wrong in production.
§4 · SELF-ASSESSMENT INSTRUMENT
Twenty questions about child safety detection that most platforms cannot answer
20 questions drawn from the framework. They are the ones platforms find hardest to answer, which is the point. You will get a readiness report showing how much of a full disclosure you could complete today, which gaps matter most, and which regulatory deadlines apply to you.
You will not get a level. Levels are earned by publishing a complete disclosure, not by filling in a form.
Nothing you enter leaves your browser. There is no account, no email required, and no submission. The assessment runs entirely on your device and we receive nothing. You can verify this in your browser's network tab.
20 questions, about 10 minutes. No account, no email, no submission. Nothing you enter leaves your browser.
§5 · INTENDED READERSHIP
Who this document is written for
Trust and safety teams
To find out what you can and cannot currently evidence, before someone else asks.
Policy and legal teams
To map detection practice against the OSA, DSA, GDPR, the Children's Code and the AI Act in one pass, with the deadlines attached.
Procurement teams
To ask vendors the questions in P2 and P5. Any vendor that cannot answer them is asking you to take detection performance on trust.
Regulators
To tell us whether this supports your enforcement objectives or cuts across them. That question is genuinely open and the comment address is below.
Researchers
Challenge the criteria, propose better ones, or tell us where the statistics are wrong.
Children and young people
The framework asks platforms what happens to a young person when detection gets it wrong. We want to know whether we asked the right things.
§6 · DECLARATION OF INTEREST
The authors are not a neutral party
This framework is published by Tuteliq AB, which sells child safety detection technology. Prof. Sarah Kingston is a paid advisor to Tuteliq. Both relationships are declared in the framework.
That is a real problem for a document like this, and we would rather state it plainly than have it discovered. Standards credibility requires separating authorship from commercial interest, and we do not have that separation. It is why this is called a framework and not a standard, why nothing here is a certification, and why we have proposed transferring stewardship to a neutral body, a BSI committee, an IEEE working group, a multi-stakeholder consortium, or an academic centre, during 2027 to 2028, with Tuteliq as one contributor among others.
- The assessment does not sell anything. No email gate, no demo request, no contact form on the results page.
- We do not receive your answers. The assessment runs in your browser and transmits nothing.
- We are publishing our own disclosure against our own framework, including the categories where we perform badly.
If you think the framework is shaped around what we happen to sell, tell us which clause and we will publish the criticism alongside the response.
§7 · COMMENT AND REVISION
This document will change
This is a working draft published for comment, not a finished instrument. We expect parts of it to be wrong: legal references we have misread, thresholds that do not survive contact with production systems, questions that turn out to be unanswerable as written.
Tell us which ones. A misstated statutory reference, a statistical claim that does not hold, a requirement that is impractical at the scale you operate at, a question we should have asked and did not. Corrections are made in the next version and recorded in the version history.
We are particularly interested in hearing from anyone who tries to complete a disclosure and finds they cannot. That is more useful to us than agreement.
Comment period closes 31 December 2026. research@tuteliq.ai
§8 · FREQUENTLY ASKED QUESTIONS
Questions raised during drafting
Is this a certification?
Do you store our answers?
Why is a detection vendor publishing this?
Isn't the top tier just a description of your product?
What if we score badly?
Can we put a badge on our site?
Does this replace OSA or DSA compliance?
Our detection is a third-party vendor's. Who answers?
Who is this for?
Levels measure disclosure completeness, not detection quality. A platform with excellent disclosure of weak detection scores higher than one with strong detection and poor documentation. Levels are not comparable across platforms as evidence of quality.
FULL DOCUMENT
Download TUT-CSDT-001 v1.0 as a PDF
The complete framework, including the six pillars, the 20-question instrument and the scoring rules. We ask for your work details so we know which organisations are reviewing the draft during the comment period.
Download the guide
Complete the form to receive the PDF instantly. Work email required.
EU-hosted CRM. No third-party ad tracking on this form.
HOW TO CITE THIS DOCUMENT
Tuteliq AB (August 2026). Child Safety Detection Transparency Framework, TUT-CSDT-001 v1.0. Working draft for public comment. CC BY 4.0. https://tuteliq.ai/standard
Licensed under CC BY 4.0. You may reproduce, translate, and adapt this framework with attribution, including for commercial purposes. Nothing in this document constitutes legal advice.