Solitude — a figure before the immensity (Friedrich): the quiet that surrounds a child set apart
Research journal — open, honest, evidence levels shown

The “murmur”: what real interaction data tells us about veiled bullying — and what it doesn’t

Bullying between minors is not in a single message, but in the deformation of a relationship over time. This journal shares, at each step, what we find — including, and especially, our failures. Figures are dated, evidence levels are shown, limits are written plainly.

Issue No. 1 — 10 June 2026Status: exploratorySOS ÉCRANS · non-profit
The bottom line
On simulated data we define exclusion by removing the very interactions a detector measures — so “detection” is partly circular. A synthetic bench can establish false positives, sensitivity and delay — never detection itself. Real validation needs labelled field data. The bottleneck is not the model — it is the data.
This issue

What we tested this week

We tested detectors of relational deformation on a real interaction network (Copenhagen Networks Study — ~700 students, 4 weeks, no content: Bluetooth proximity + timestamped sms/calls), by simulating a progressive exclusion of one target, and measuring — honestly — false positives, sensitivity and delay. ExploratorySimulated · adults

0%
false positives
(clean baseline)
~30%
network detectors
(detection plateau)
48%
“unrepaired rupture”
best signal

The figure that matters

Detection vs. how stark the exclusion is
0% 30% 60% soft exclusion stays invisible none soft partial total 57%
“Unrepaired rupture” detector, behavioural eviction. Detection climbs only as exclusion approaches total; gradual, deniable exclusion — the actual murmur — stays invisible. 0% false positives throughout.

Three findings, no spin

1 · Network detectors only see brutal exclusion

Early-warning (AR1), k-core, irreversibility and sub-network detectors plateau at ~25–30% detection, at 0% false positives. Soft, gradual exclusion stays invisible.

2 · The clearest signal: the “unrepaired rupture”

The share of a target’s messages that go unanswered detects ~48% of exclusions at 0% false positives — but it is partly circular (we remove replies; the feature counts replies). What stands: 0% false positives, and soft exclusion still missed.

3 · Channels are decoupled (a circularity test)

Exclusion injected into physical proximity is not visible in replies, and vice-versa: a detector does not magically “guess” another channel. You must instrument the channel where exclusion actually happens.

What this means for the tool. Since no simple detector sees the veiled murmur, we refuse to promise detection we cannot deliver. We shift the value: not “the app that detects”, but a companion that helps the child self-observe (gentle, on-device, nothing leaves), gives them words to name what they live, and keeps a human in the loop — the tool only flags a strong rupture, “for an adult to look at”, never a verdict.

What we seek — and what we offer

We offer these results and our scripts to anyone. We seek: (1) real labelled peer-interaction data (class climate), or a field collaboration (a school pilot is being prepared for the new term); (2) failing that, simply your critical eye on the method. No reply expected — this is a gift, not a request.

Limits we own

Simulated eviction on adults: exploratory, not proof on children.
Wide confidence intervals (~30 targets): nothing is statistically separated.
“Detection %” on simulated data is partly tautological (see the bottom line).
Independent validation is intended, not yet done.
📬 Following this journal. It updates at each step. If a brief note at each issue is useful to you, reply once — and you can unsubscribe at any time, one click. Otherwise this page stays here, at your own pace.