Spotting the school bullying that doesn't shout — quiet exclusion, the deniable jab — without ever reading a single message.
This is not a funding request: it is an invitation to a critical look. I carry this idea as a researcher, not a salesman. The project is non-profit and free. If the angle speaks to you, let's talk; if not, your objections are already valuable to me.
The full conversation carries a signal that an isolated message does not: on the same corpus, a detector goes from indistinguishable-from-chance at the message level (AUC 0.556) to a clearly informative score on the full thread (0.762). It remains to be proven that this gain comes from relational structure — not merely from more context. That is precisely the collaboration I am looking for.
Does relational aggression leave a measurable signature in the dynamics of a relationship, independently of message content? If so, school bullying is only one application among others (exclusion, isolation, manipulation). It is this question — the observation of a phenomenon, not the performance of a product — that we submit for your critique.
Around 200 volunteers · over 6,000 judgements · 12 days of collection. The facts, unembellished.
| Measure | Value | Honest reading |
|---|---|---|
| Inter-rater agreement (Krippendorff α, 9 raters) | 0.574 | Below our own floor (0.667). Even human experts struggle to agree on what counts as veiled bullying — the question is first of all one of definition. |
| AUC, message → thread (n=39) | 0.556 → 0.762 | Gain of +0.21, suggestive (no DeLong test). May reflect "more context": proving it is relational is the point of the pilot. |
| Recall on the murmur (the target) | 50-60% | Half of subtle cases still escape. Innovation not yet demonstrated — the priority. |
| False positives on real data (thread level) | 5.40% | Too high for a minor → distribution suspended. |
| Positive predictive value (derived, 2% prevalence) | ~18% | About 4 alerts in 5 would be false. A decisive figure, not yet measured on the target. |
A solid concept but not validated on the target population (ages 8-13): all results are on adolescents. At low prevalence, even a good detector produces a majority of false alerts — hence an aggregated signal, never an accusation, and the estimation of real predictive value as the number-one deliverable of a pilot.
The decisive lock: the literature offers no baseline at this age. Without it, no claim about the target is possible.
Reconciling the child's right to privacy (UNCRC art. 16) with parental authority, the digital age of consent being 15 in France. To be worked out with a lawyer — not a dogma on our part.
With a child psychiatrist: modelling the impact of a false-alert rate on a child before any calibration.
No money or funding · no deployment · no commercial access · no reading of children's content. Local processing is still processing under the GDPR — which is exactly lock no. 2.
A thirty-minute conversation, no commitment. The full scientific dossier — detailed results, method, register of discarded leads, 184 references — awaits you.
Read the full dossier