Last week we introduced the Context Knowledge Base — the structured semantic layer behind PurenetX. Today we go deeper: what's actually inside it, why it's built the way it is, and why we believe knowledge is the hard dependency of AI safety that almost nobody talks about.
Here is the uncomfortable truth about most visual safety systems: they are exercises in threshold-tuning. Detect skin above X%, proximity below Y pixels, flag. When the false positives pile up, nudge the threshold. When something harmful slips through, nudge it back. Teams spend years in this loop, and the system never once understands what it is looking at.
Our bet is different, and it is unglamorous: a model that knows more about the world makes better safety decisions than a model that just sees more pixels. You can't threshold your way to understanding.
The problem: thresholds don't know anything
Pixel-based systems fail in both directions, and both failures trace to the same root.
False positives. A dermatology exam gets flagged as suspicious content. A wedding embrace gets flagged as exploitation. A gym session trips the exposure threshold. Legitimate content is blocked, users lose trust, and creators leave.
False negatives. A covert recording — visually unremarkable, low exposure, nothing crossing any threshold — passes untouched. The genuinely harmful case often looks the safest to a pixel detector.
The root cause is that thresholds carry no model of setting, role, or culture. Consider the same medical examination performed in a clinic in Chennai and one in Chicago. Different room layouts, different attire norms, different equipment, different visual signatures — the same legitimate procedure. A threshold tuned on one context misfires on the other. Scale that across every culture a global platform serves, and threshold-tuning becomes a game you cannot win.
What's missing is not sensitivity. It's knowledge.
Inside the CKB
The PurenetX Context Knowledge Base (CKB) is our answer: a professionally structured collection of visual safety scenarios, currently 55+ scenarios and growing weekly, built to teach a system what's happening, not just what's visible.
It is organized into five categories, chosen because they cover the highest-stakes decision boundaries in visual safety:
- Medical — clinical contexts that pixel systems chronically misflag
- Fitness — athletic contexts with high visual ambiguity
- Celebration — social contexts where physical closeness is expected and consensual
- Intimate — private contexts where consent is the deciding line
- Safety — scenarios containing genuine harm signals that thresholds miss
Each scenario is annotated across our six context dimensions — scene, activity, relationship, intent, platform, and risk — expanded into 13 annotation fields per scenario: scene context, activity type, relationship context, intent signals, consent visibility, vulnerability assessment, risk indicators, ground-truth decision, decision reasoning, and annotator confidence, among others.
Two fields deserve special mention. The ground-truth decision and its reasoning are part of the data itself — every scenario records not just what the right call is, but why. That is what makes explainable output possible downstream: a system trained on explicit reasoning can produce reasoning of its own.
On confidence: CKB v1 annotator confidence currently ranges from 0.79 to 0.94 across scenarios. To be precise about what that number is — it reflects how confident human annotators are in each scenario's labels. It is not a model performance metric. Model performance numbers are pending peer review, and we won't publish them before that.
One example: the wedding
The celebration category shows the difference most sharply.
A wedding embrace: physical contact, close positioning, emotional intensity. A pixel-based system reads it as skin contact + intimate positioning → BLOCK. Another false positive; another user wrongly punished for a celebration.
The CKB-informed reading: celebration setting detected, ceremonial activity, relationship context inferred from the scene (not tracked across sessions — nothing about the individuals is stored), consent indicators present → ALLOW.
Same visual signal. Different decision. Better outcome — and crucially, an explainable one. The system can state why it allowed the moment, in terms a trust-and-safety team can audit.
That is the whole thesis in one scenario: the pixels were identical; the knowledge made the difference.
The unglamorous truth
Building the CKB is slow, methodical work, and we want to be honest about that.
Every scenario requires structured thinking across cultures, settings, and professional domains. Contributing researchers typically invest two to three hours per scenario set — defining the context, debating the boundary cases, documenting the reasoning. Annotations are reviewed by multiple people, disagreements are resolved through written guidelines, and the whole thing is versioned like code.
Nobody writes headlines about annotation guidelines. But this is the backbone: a safety system only works in the real world if it understands the real world, and that understanding has to be built by hand before it can ever be learned by a machine.
What's next
We are expanding the CKB through research partnerships. As of this week, we've reached out to 14 researchers, and we expect 3–5 active partnerships by the end of July. Contributors are credited, and the work is heading toward co-authored, peer-reviewed publication — paper submission targeted for Q3 2026, publication targeting Q2 2027 at venues in the FAccT / CHI / USENIX orbit.
On access: representative public scenarios from the medical, fitness, and celebration categories are available on request, so researchers can evaluate the annotation schema directly. The full CKB is available to research partners under NDA — a deliberate choice that protects both the sensitive nature of the safety and intimate categories and the integrity of future benchmark evaluations.
In parallel, the CKB feeds our design partner pilots with early-stage camera platforms — real-world scenarios from pilots expand the CKB, and the CKB improves pilot decisions. That loop is the product.
The unsexy conclusion
The truth about AI safety that doesn't fit in a headline: understanding matters more than threshold-tuning. More pixels, higher resolution, bigger detection models — none of it closes the gap between what's visible and what's happening. Knowledge closes that gap.
We're building the knowledge layer. It's not flashy. But it works — and we'd rather build the unglamorous thing that works than the impressive thing that doesn't.
Want to dig in? Request access to the public CKB scenarios, read about our research direction, or reach out about a research partnership — we review every serious inquiry weekly.
— Kaushik A., Founder & CEO, PurenetX