Governance
Authority through independence
Authority through independence
The people who define what harm looks like are not the people who sell the system.
Nostra Veridicus is researching and developing a safeguarding intelligence layer that platforms integrate into their existing infrastructure. It is a commercial product, sold to platforms as a service. What out guardian style layer will do is different from every content moderation tool on the market. Rather than making a binary decision about a single piece of content, it monitors anonymised behavioural and linguistic signals over time, building a longitudinal picture of how a user's patterns are changing, and whether those changes suggest escalating risk. It then generates a recommendation to the platform about what, if anything, to do next.
The platform acts. Our integration advises. The decision and the action always rest with the platform. Not with us.
Our guardian style layer does not remove content, or contact users directly. It analyses patterns, assesses risk, and recommends a proportionate response. Every recommendation above the lowest level is reviewed by a trained human moderator before it reaches the platform.
How the escalation framework works
Risk is not binary. A person whose signals are shifting in a concerning direction is not in the same situation as someone in immediate danger, and the response should not be the same either. We operate a five-level escalation model, designed so that the intervention is always proportionate to the assessed risk.
| Level | Name | Signal profile | What the platform receives | Human review |
|---|---|---|---|---|
| 0 | Baseline | Normal activity. System establishes individual behavioural and linguistic baseline over an initial observation window. | Nothing. Passive monitoring continues. | None |
| 1 | Early signal | Isolated markers in one category. Below pattern threshold. Logged and monitored for development. | Low-key adjustment to the mix of content shown, with more positive, community, and wellbeing material. Delivered as a platform-wide wellbeing feature that is disclosed to all users in plain language. No individual is told they were flagged. Design under research and Board review. | Logged. Reviewed in weekly batch. |
| 2 | Emerging pattern | Repeated signals across multiple sessions over days or weeks. Pattern developing. Possible single-category escalation. | Platform surfaces opt-in wellbeing resources as a general feature, not targeted at the user. NV moderator notified. | Required before recommendation is issued. |
| 3 | Elevated concern | Consistent escalating pattern. Multiple signal categories co-occurring. Sustained over time. Possible victim or perpetrator indicators. | Moderator reviews full signal history. Platform may surface location-specific support service information. Educational platform safeguarding lead alerted. | Mandatory. Senior moderator sign-off. |
| 4 | Imminent risk | Explicit signals of imminent self-harm, harm to others, or trafficking and exploitation. Acute escalation velocity. Confirmed by two-moderator review. | Immediate escalation to platform designated safeguarding lead. Recommendation to contact relevant authorities. Emergency SLA applies. | Mandatory. Two-moderator sign-off required. |
- Signal profile
- Normal activity. System establishes individual behavioural and linguistic baseline over an initial observation window.
- Platform receives
- Nothing. Passive monitoring continues.
- Human review
- None
- Signal profile
- Isolated markers in one category. Below pattern threshold. Logged and monitored for development.
- Platform receives
- Low-key adjustment to the mix of content shown, with more positive, community, and wellbeing material. Delivered as a platform-wide wellbeing feature that is disclosed to all users in plain language. No individual is told they were flagged. Design under research and Board review.
- Human review
- Logged. Reviewed in weekly batch.
- Signal profile
- Repeated signals across multiple sessions over days or weeks. Pattern developing. Possible single-category escalation.
- Platform receives
- Platform surfaces opt-in wellbeing resources as a general feature, not targeted at the user. NV moderator notified.
- Human review
- Required before recommendation is issued.
- Signal profile
- Consistent escalating pattern. Multiple signal categories co-occurring. Sustained over time. Possible victim or perpetrator indicators.
- Platform receives
- Moderator reviews full signal history. Platform may surface location-specific support service information. Educational platform safeguarding lead alerted.
- Human review
- Mandatory. Senior moderator sign-off.
- Signal profile
- Explicit signals of imminent self-harm, harm to others, or trafficking and exploitation. Acute escalation velocity. Confirmed by two-moderator review.
- Platform receives
- Immediate escalation to platform designated safeguarding lead. Recommendation to contact relevant authorities. Emergency SLA applies.
- Human review
- Mandatory. Two-moderator sign-off required.
The most important design choice in this framework is what happens at Level 1. The first response to an emerging signal is deliberately gentle: a low-key adjustment to the content a user sees, increasing the presence of positive, community-oriented, and wellbeing-related material. The platform discloses this in plain language as a general feature that applies to everyone, so no one is influenced without being able to find out how. No individual is told they have been flagged, and no action is taken against them. The signal is logged and reviewed in a weekly batch, so that false flags can be identified and thresholds corrected. If the signal was a false flag, the user has seen a slightly different mix of content and nothing more. This is proportionality in practice, not in principle.
We do not assume this approach is harmless or effective. Whether a gentle, disclosed adjustment improves outcomes, and whether it can affect people who were never at risk, are open research questions. The Level 1 design is part of our research programme (see Theme 4 on the Research Opportunities page), is subject to review by the Expert Advisory Board, and may change as evidence emerges.
The governance question this creates
Building a system that influences what people see, recommends platform interventions, and advises on escalation to emergency services creates an immediate and serious question: who decides where the thresholds are, and on what basis?
This is not a theoretical concern. A threshold set too low generates false positives that cause harm, particularly to communities whose communication styles are already over-policed. A threshold set too high misses people who need support. A marker that is clinically valid but culturally specific will perform differently across different user populations. These are consequential decisions.
If those decisions are made by the commercial team at Nostra Veridicus, they will be made with commercial interests in the room. That is not acceptable for a system designed to protect vulnerable people. So we removed the commercial team from that decision entirely.
The Expert Advisory Board
Expert Advisory Boards will hold formal authority over the harm signal taxonomy, the escalation thresholds, and the minimum acceptable model performance standards. The commercial leadership of Nostra Veridicus does not hold a deciding vote on any of these matters. This is written into the governance charter and disclosed contractually to every platform partner.
The board is constituted across six domains. Each contributes something the others cannot, and the absence of any one would compromise the validity of the system.
| Domain | Authority on the board | Why it cannot be absent |
|---|---|---|
| Clinical psychology and mental health | Defines the clinical validity of self-harm, distress, and coercive control markers. Sets the evidence standard for what constitutes a meaningful signal. | Without clinical grounding, markers are proxies, not evidence. Proxies cause harm. |
| Legal and human rights | Reviews all markers and protocols for compliance with the ECHR, UK GDPR, the Children's Code, and the Online Safety Act. Advises on subject access response design. | A system that cannot withstand legal scrutiny cannot be deployed. A system that violates human rights should not be. |
| Survivor and lived experience | Reviews markers for alignment with how harm is actually experienced, not how it appears to professional observers. Multiple representatives across harm categories. | Clinical abstraction can miss what lived reality makes obvious. The people most affected must have a voice in how the system works. |
| Elevated concern | Consistent escalating pattern. Multiple signal categories co-occurring. Sustained over time. Possible victim or perpetrator indicators. | Moderator reviews full signal history. Platform may surface location-specific support service information. Educational platform safeguarding lead alerted. |
| Imminent risk | Explicit signals of imminent self-harm, harm to others, or trafficking and exploitation. Acute escalation velocity. Confirmed by two-moderator review. | Immediate escalation to platform designated safeguarding lead. Recommendation to contact relevant authorities. Emergency SLA applies. |
How the board operates
- The board will meet regularly. Between sessions, written reports on model performance, false positive rates by demographic group, and proposed taxonomy changes are circulated for review.
- Any addition, amendment, or removal of a harm marker requires board approval by supermajority, with written rationale and supporting evidence submitted in advance.
- An independent external audit of board governance will be conducted annually. Summary findings are published.
- A formal whistleblowing mechanism exists for board members to raise concerns about commercial pressure on research or governance decisions. This mechanism will be administered externally to Nostra Veridicus.
- Board members are remunerated at rates commensurate with their professional standing. Governance cannot be meaningful if it depends on unpaid goodwill.
Scope Boundary
The board also holds authority over the scope of the system, specifically, what it will never monitor. This boundary is written into the governance charter, included in every platform contract, and cannot be overridden by any commercial agreement.
- Self-harm and suicidal ideation
- Grooming and child sexual exploitation
- Coercive control and psychological manipulation
- Violence, targeted threats and escalating aggression
- Emotional and psychological abuse
- Human trafficking and exploitation indicators
- Predatory behaviour toward vulnerable users
- Political speech, protest, or dissent
- Religious or ideological expression
- Journalism or whistleblowing
- Content that is offensive or controversial
- Government or law enforcement surveillance
- Commercial or reputational monitoring
- Anything outside the validated harm taxonomy
Law enforcement
Nostra Veridicus is not an instrument of government, law enforcement, or any state body. We are a third-party supplier to platforms, operating as an independent commercial entity.
Law enforcement engagement is limited strictly to the imminent harm escalation pathway (Level 4 in our five-level model), where there is credible, human-reviewed evidence of risk to life or safety. At this level, the recommendation to the platform may include contacting relevant authorities. The decision to act on that recommendation rests with the platform and its designated safeguarding lead, not with Nostra Veridicus.
We will engage in early-stage conversations with cyber and counter-exploitation law enforcement contacts to understand operational requirements at this pathway. Those conversations inform the design of the imminent harm escalation protocol. They do not give law enforcement any access to our systems, data, or processes.
Any platform requesting expansion of scope beyond the validated harm taxonomy will be declined. This is not a negotiating position. It is a governance constraint.
Ready to partner with us?
We work with organisations that share our commitment to evidence-based practice.