Our Approach

Phase I

Theoretical grounding

Before any data is collected or any online space is observed, the research must establish its theoretical foundations.

This phase produces three things:

• A systematic literature review spanning the clinical psychology of harm, coercive control theory, developmental psychology of adolescent vulnerability, criminological frameworks for predatory behaviour online, and the linguistics of manipulation and distress. The review is conducted to academic standard, comprehensive, critically appraised, and documented.
• A harm signal taxonomy, a structured map of the categories of harm the system is designed to detect, with each category grounded in at least two independent peer-reviewed sources. The taxonomy is the foundational document of the entire project. It defines what we look for, why we look for it, and what the evidence base is.
• An academic ethics application, covering research methodology, data handling, participant protections, and the basis for observational online research. No data collection of any kind begins before ethics approval is in place.

The taxonomy will be published in summary form. This is a commitment to transparency: the markers we use to identify potential harm should be explicable, and where possible auditable, by the people affected by them.

Phase II

Online space research

Theory must be tested against reality. The language and behaviour patterns described in clinical literature may manifest differently, or not at all, in gaming chat environments, educational platforms, and the subcultural spaces that adolescents and young adults actually inhabit online.

This phase conducts systematic observational research in target platform environments under ethics approval. The questions are:

• Are the theoretically derived harm signal markers present in these environments? In what form and with what frequency?
• How do platform-specific contexts shape the expression of those markers? The language of coercive control in a gaming clan looks different from the language of coercive control in an educational messaging tool.
• What does baseline, healthy interaction look like in these spaces? Defining the normal is as important as defining the signal, without it, any deviation from baseline looks like a marker.
• What signals exist that theory has not yet captured? Practitioner knowledge and careful observation often surface patterns that academic literature has not yet formalised.

This phase also begins stress-testing markers for false positive risk, particularly for communities whose communication styles are frequently and incorrectly flagged as threatening by systems calibrated on narrow data.

Phase III

Dataset construction

The dataset is the core deliverable of the research programme. It will be constructed to a standard that makes it usable for training detection models, publishable as an academic resource, and defensible under clinical and legal scrutiny.

Construction principles:

• All data is collected under ethics approval, with appropriate consent frameworks where applicable.
• Multi-category annotation. Each example tagged across harm type, escalation level, confidence grade, cultural context, and platform type
• Diverse annotation panel. Annotators drawn from varied demographic, cultural, and professional backgrounds, with inter-annotator agreement measured and reported transparently
• Negative examples form a significant portion of the dataset, non-harm examples that superficially resemble harm signals, to train the system against false positives
• Versioned releases. The dataset is maintained in versioned releases, with each version reviewed and approved by the Expert Advisory Board before use in model training

The dataset is not sold or licensed to third parties. It is jointly owned by Nostra Veridicus and the academic partner, and its use is governed by the ethics framework under which it was constructed.

Phase IV

Expert validation

No dataset, marker, or detection model is deployed until it has been reviewed and approved by the Nostra Veridicus Expert Advisory Board.

Validation is not a one-time gate. It is an ongoing process.

Board validation covers five dimensions:

• Clinical validity: do the markers reflect current clinical understanding of how harm presents, across different age groups and circumstances?
• Cultural validity: do the markers hold across diverse communities, or do they embed bias that disadvantages particular groups?
• Legal defensibility: are the markers and escalation thresholds defensible under applicable human rights and data protection law?
• Lived experience alignment: do the markers reflect how harm is actually experienced by those who have lived it, or only how it appears to outside observers?
• False positive distribution: what is the rate of incorrect escalation across different demographic groups, and does it meet the board's minimum acceptable standard?

The board meets quarterly and reviews model performance data on an ongoing basis. Markers can be retired, revised, or added based on emerging research or observed system performance. The commercial team does not hold a deciding vote on any of these matters.