In our last post, we argued that most AI vision systems fail because they see pixels, not meaning. The obvious question follows: how do you teach a system meaning?
Our answer is not a bigger model or a bigger dataset. It is a different kind of dataset.
Why datasets aren't enough
Modern vision datasets are remarkable at answering two questions:
- What objects are present? — person, bed, phone, medical equipment
- What actions are visible? — stretching, embracing, recording
But they rarely answer the questions that safety decisions actually depend on:
- What is actually happening here?
- Is the activity consensual?
- Is this professional, medical, educational — or harmful?
You can scale an object-detection dataset to a billion images and it will still not encode the difference between a clinical examination and a violation. The label vocabulary itself has no room for that distinction. This is not a data-volume problem. It is a data-structure problem.
What is a Context Knowledge Base?
The PurenetX Context Knowledge Base (CKB) is our response. A CKB does not replace visual data — it adds structured context around it. Every scenario in the CKB is annotated not just with what is visible, but with what it means:
| Field | Example |
|---|---|
| Scene | Wedding |
| Activity | Celebration |
| Relationship | Couple |
| Intent | Ceremony |
| Consent | Implied |
| Decision | Allow |
| Reason | Celebration context |
The decision and its reason are part of the data. That is what makes explainable output possible downstream: the system is trained on scenarios where the why is explicit, so it can produce a why of its own.
Why context matters: three scenarios
Consider three scenarios that pixel-level analysis struggles to tell apart:
- Wedding celebration — physical closeness, emotional intensity → Allow. Event context and celebratory intent are clear.
- Medical examination — exposure in a clinical setting → Allow. Professional context, clinical relationship, procedural intent.
- Secret recording — often visually unremarkable → Block. Covert-capture indicators, absent consent.
Pixels alone cannot distinguish these. In fact, by pixel measures the harmful case often looks the safest. Only the contextual layer — relationship, intent, consent — separates them. The CKB is built precisely around such contrast sets: scenarios that look similar but mean different things.
Human annotation, done seriously
Context is a human judgment before it can ever be a machine one. So the CKB begins with human annotation — and we treat that as research infrastructure, not a labeling chore:
- Multiple reviewers per scenario, because a single annotator's cultural lens is a bias vector.
- Written annotation guidelines that define each field, each value, and the boundary cases.
- Disagreement resolution — when reviewers disagree, that disagreement is data. It flags genuinely ambiguous scenarios, which are exactly the ones a safety system must handle honestly.
- Version control — the CKB is versioned like code. Guidelines evolve, annotations are re-audited, and every model evaluation records which CKB version it ran against.
Diversity is a design requirement throughout: scenarios span cultural contexts, demographics, and skin tones, so that fairness can be measured rather than assumed.
An evolving research direction
The CKB is deliberately unfinished. Coverage grows, the taxonomy gets refined, and annotation quality improves with every review cycle. To push it further we are actively seeking:
- Researchers interested in context representation, annotation methodology, and fairness evaluation
- Universities for independent validation and co-authorship on the benchmark work ahead
- Design partners whose real-world scenarios expand coverage where it matters most
If that's you, get in touch — or read more about our research direction.
Conclusion
Context intelligence begins with better understanding, not larger models. Building that understanding requires structured knowledge alongside visual data — a semantic layer that captures relationship, intent, and consent, annotated carefully by humans and versioned like the research asset it is.
That is what the CKB is. And it is the foundation everything else we build stands on.
— Kaushik A., Founder & CEO, PurenetX