AUN Coding Hub · Google Africa Applied AI Lab 2026

Making absence visible.

Digital Roundabouts for Peace — a working system that shows which stories dominate a community's shared spaces, and whose stories go quietest.

Manara Coexistence Framework · American University of Nigeria — live at dejm.digitalriver.ng

01 · Origin

A roundabout in Mubi North

In Mubi North, Adamawa State, a physical roundabout bears the symbols of the Gude, Fali, and Fulani ethnic groups — a piece of infrastructure that compels everyone passing through it to acknowledge their coexistence. This project asks what the digital equivalent would look like, because contemporary conflict doesn't announce itself in the street. It accumulates quietly, inside the family WhatsApp groups and religious networks that we call Trusted Digital Micro-Publics — spaces where a story told by a respected elder is never questioned, and where silence functions as assent.

The research behind DEJM is grounded in 8 in-depth interviews and 3 ethnicity-specific workshops conducted in Mubi North, Adamawa State, Nigeria.

A family shares a story: their ancestors migrated from "Far Cameroon" generations ago. The story is accurate. It is told by respected elders. No one disagrees.

Over time it becomes the lens through which political legitimacy is assessed. When a candidate emerges whose lineage lacks that link, the question arises — not in rallies or debates, but in trusted circles:

"Can someone without our history truly represent us?"

From the underlying research paper — the narrative never becomes hate speech. It is never fact-checked, because it is factually sound. It circulates through inheritance, not propaganda.

02 · The Mechanism

The Tacit Score Index

TSI quantifies how much a historical narrative is being tacitly weaponized — without ever asking whether the narrative itself is true.

Cw×Ts×Ar = TSI
Cw Centrality Weight

How foundational the narrative is to community identity — origin, belonging, lineage, or political legitimacy.

Ts Source Trust

Epistemic authority of who is telling it. Community-defined trust tiers — elders score highest, anonymous senders lowest.

Ar Affective Resonance

Scored inversely to antagonism. The quieter and more naturalised a narrative, the higher it scores — and the more consequential it's treated.

That inversion is the paper's central claim: a loud, hostile message is easy to spot and easy to dismiss. A calm, uncontested, generations-old story is the one nobody thinks to question — and the one doing the most work.

Benign · <200 Medium Risk · 200–499 High Risk · ≥500 Archived Silence · separate signal

Archived Silence isn't a TSI band — it's a distinct check: when a narrative our domain experts flag as high-potential (high centrality, high trust) circulates nowhere at all in a given period. The dashboard surfaces it with a fixed, reflective prompt: "Why are some foundational stories never told in our shared conversations?"

03 · A Working System

Not a framework diagram

Every stage below is running code, not a proposal — a message moves through all six before anything is scored.

  1. Ingest. WhatsApp / Telegram export upload, or a live Telegram Bot webhook.
  2. Consent gate. A message from anyone without active, per-group consent is discarded before it's ever stored — encrypted or not.
  3. Strip & encrypt. Phone numbers, emails, handles, and titled names removed; what remains is AES-256 encrypted.
  4. Score. Every 5 minutes, queued batches are analysed and scored — Cw / Ts / Ar computed, TSI assigned a risk band.
  5. Detect. Archived Silence checked against expert-seeded narrative baselines for the period.
  6. Alert. Hourly governance check — a community is notified when one group's narratives cross its own configured dominance threshold.
Non-adjudicative by rule, not just by policy. The scoring model is explicitly instructed to never use the words "true," "false," "propaganda," "hate speech," or "misinformation," and never to recommend suppression, removal, or moderation of anything. It analyses how narratives circulate — not whether they should.
04 · Validated, Now Ready for the Field

Phase 1 is done. Phase 2 is the ask.

Phase 1Done

Simulation-validated

You don't point an unvalidated conflict-monitoring tool at a live, vulnerable community before proving it works. So the full pipeline — ingestion through consent, scoring, silence detection, and governance alerting — was proven end to end against realistic synthetic datasets, including one deliberately modeled on Mubi North's own Gude / Fali / Fulani / displacement narrative themes.

39,820messages processed
897TSI computations
1,657audited, chain-verified actions
0raw messages retained after scoring
05 · Inference

Claude, bound by rule

Every batch is analysed by Claude Sonnet (claude-sonnet-4-6), given the anonymised, tokenised messages alongside the community's own expert-seeded trust tiers and narrative baselines — never a generic prompt with no local context.

Today, inference is called directly — a deliberate, honest architectural choice at this stage, not an oversight. It means the fastest path to adding a second engine, such as Gemini, is a contained change: a second implementation of the same structured request-and-response shape the pipeline already runs on, not a rebuild.

06 · Safeguarding

Consent first, memory last

Monitoring community discourse in a conflict-affected state carries real ethical weight. The system is built to protect participants from itself.

07 · Team

Five researchers, one university

Dr. Zainab Shamsiya UsmanFaculty Advisor · Software Engineering
Alvin Orisedavweji FortetaInformation Systems
Muhammad Dauda BelloInformation Systems
Kamal Jillahi BakariComputer Science
Joseph Abunimye IngioComputer Science

School of Information Technology and Computing, American University of Nigeria.

08 · The Ask

What would move this into the field

01

Model access

API credits and access to Gemini, to build and validate it as a second, interchangeable scoring engine alongside Claude.

02

Engineering mentorship

Hardening the pipeline for real field conditions — unreliable networks, consent flows at scale, and the edge cases a validation run doesn't surface.

03

Field deployment funding

Resources to run Phase 2 itself: live, consent-gated ingestion from real Mubi North Trusted Digital Micro-Publics.

"Sustainable coexistence requires the deliberate engineering of epistemic environments that make the forgetting of plurality impossible."