// Multi-Location Multi-Agent Research Drill — Nigerian Scenario Library

Integrated Research Drill

⚠ Published metrics preserved — simulation only 95.3% accuracy · 127 ms latency · 10,847 windows 35 Nigerian locations · 3 risk corridors

Simulation Parameters

Range 1–2,000
Range 1–100
200
20%
Location Scenario Library — select up to 5 simultaneous locations
Zamfara State
Katsina State
Kebbi State
Sokoto State
Kaduna State
Niger State
Kogi State
Kwara State
Plateau State
Benue State
Nasarawa State
Borno State
Yobe State
Adamawa State
Bauchi State
Taraba State
FCT / South-West
South-South
South-East
North-West Urban
IDLE
// Phase 01 — Urban Mobility Edge Intelligence Test

Abuja Metropolitan Corridor

Validates sub-150 ms alert latency and network failover across the configured device population. 5G collapse at t=45s stress-tests the adaptive FSM.

100 devices 5 schools CNN-LSTM INT8 · 22ms inference 5G → WiFi → LoRaWAN FSM

Research objective

Demonstrate sub-150 ms latency and <3s network recovery under 12-second 5G blackout. Gateway utilisation modelled dynamically from configured device count and gateway capacity.

GATEWAY
Jetson Nano
TensorRT · 472 GFLOPS
WEARABLE
ESP32-C3
RISC-V 160MHz · 5µA sleep
MODEL
CNN-LSTM INT8
1.07 MB · 22ms · 95.3%
GW CAPACITY
200 devices
Load-balanced if exceeded
HYSTERESIS
2-window filter
Prevents false positives
DISRUPTION
12s 5G blackout
Forced at t=45s
LATENCY (ms)
NET SWITCH (s)
PACKET LOSS
GW UTILISATION
ceics-gateway-abuja · phase-1 · 100 dev / 5 schools
[ gateway-abuja ] ready — configure parameters then press RUN
// expected targets
Alert latency<150 ms
Network recovery<3.0 s
Packet loss<5%
GW utilisation<80%
// observed results
Alert latencypending
Network recoverypending
Packet losspending
GW utilisationpending

// latency — 5G collapse at t=45s

// Phase 02 — Rural Autonomous Intelligence Test

Biu, Borno State

Tests edge autonomy under zero cloud connectivity across the configured population. Federated learning with Paillier HE across school gateways.

100 devices 3 km dead zone Paillier 2048-bit HE Dirichlet α=0.5

Research objective

Demonstrate 95.3% anomaly detection accuracy without cloud dependency. FL convergence within 20 rounds under non-IID data distribution across 5 school gateways.

DEAD ZONE
3 km corridor
Zero cellular · LoRaWAN only
FL NODES
5 gateways
Non-IID Dirichlet α=0.5
ENCRYPTION
Paillier 2048-bit
112-bit classical security
LOCAL EPOCHS
5 per round
SGD η=0.01
DEVICES/GW
20
Auto-distributed
ANOMALY
t=80s spike
C1+C3 co-occurrence

Federated learning — 5 school nodes

// FedAvg + Paillier HE · Algorithm 1
BROADCAST θ^(t−1) to all Gᵢ via 5G
LOCAL TRAIN 5 epochs · SGD η=0.01 · batch=32
ENCRYPT E(Δθᵢ) ← Paillier.Encrypt(Δθᵢ, pk)
AGGREGATE E(Δθ_global) ← ⊕ᵢ (nᵢ/n)·E(Δθᵢ)
DECRYPT 2-of-3 distributed key split
ceics-gateway-biu · phase-2 · 100 dev
[ gateway-biu ] ready — press RUN
// expected targets
Detection accuracy95.3%
FL convergence≤20 rounds
Energy/inference≤2.5 W
Bandwidth reduction≥60%
// observed results
Detection accuracypending
FL convergencepending
Energy/inferencepending
Bandwidth reductionpending

// FL convergence · non-IID

0.5

// energy comparison

// Phase 03 — Forest Emergency Recovery Test

Sambisa Forest Terrain

Tests store-and-forward buffering and PoA blockchain audit under zero terrestrial connectivity.

100 devices Zero terrestrial coverage Direct-to-satellite LoRaWAN PoA blockchain · 84 TPS

Research objective

Prove forensic accountability and continuous distress signalling under complete communications blackout. Validate store-and-forward buffer recovery. Blockchain audit completeness: 100%.

CONNECTIVITY
Zero terrestrial
All LoRaWAN + 5G severed
SATELLITE
Direct-to-sat
LoRaWAN · 30s beacon
BUFFER
Store-and-forward
Local cache · auto-flush
BLOCKCHAIN
PoA · 84 TPS
6.3× max alert rate
KEY SPLIT
2-of-3 multi-sig
No single key holder
AUDIT
100%
vs ~71% cloud baseline

PoA blockchain ledger — live

0x0000...genesis
GENESIS
CEICS-root
⚙ system
0x????
awaiting
unsigned
0x????
awaiting
unsigned
0x????
awaiting
unsigned
0x????
awaiting
unsigned
Confirmed: 1 TPS: 84 Completeness: 0% Consensus: PoA
ceics-satellite-relay · phase-3 · 100 dev
[ satellite-relay ] ready
// expected targets
Satellite latency<1500 ms
Audit completeness100%
Buffer flush<10s on recovery
PoA throughput≥80 TPS
// observed results
Satellite latencypending
Audit completenesspending
Buffer flushpending
PoA throughputpending

Event timeline

t+0sWearable enters Sambisa corridorNOMINAL
t+8sLoRaWAN RSSI −128 dBm — degraded modeDEGRADED
t+14sAll terrestrial severed — store-and-forward activatedBUFFER ON
t+18sTactical Recovery Module — satellite acquisitionRECOVERY
t+24sSatellite uplink established · 1.24s latencyUPLINK
t+25sBuffer flushing — 0 buffered alerts transmittingFLUSH
t+30sC2+C1 distress packet · confidence 98.7%ALERT
t+31sPoA blockchain: 4 blocks confirmed · 100% auditLOGGED
// Integrated Research Dashboard

Comparative Evaluation Report

Across configured population — all published metrics preserved from peer review submission.

CEICS validated

All 5 operational requirements jointly satisfied · Published metrics preserved

Research dimensionCEICSCloud baselineΔ improvement
Alert latency (mean)127 ms ±3.2268 ms ±8.1−52.6% · d=21.6
Detection accuracy95.3% ±0.41%88.4% ±0.61%+6.9 pp · d=16.8
Seizure recall (C2)99.1% ±0.41%+7.7 pp vs Masci 2025
Macro F1-score94.7% ±0.38%87.0% ±0.69%+7.7 pp · p<0.001
Bandwidth35% of baseline100%−65%
Audit reliability100% blockchain~71%+29 pp
Intrusion resistance78.4% ±1.8%22.1% ±2.3%+56.3 pp · d=28.2
Offline operationFull + store-and-forwardNoneComplete resilience
−52%
Latency improvement
127 ms mean · 312 ms 95th-percentile · both within clinical thresholds. ANOVA F(3,16)=847.3, p<0.001, η²=0.994.
+6.9 pp
Accuracy improvement
FL collaboration dominant driver. 10-node federation: ~21 seizure events/month vs 2.1 single-school. Non-IID Dirichlet α=0.5.
0 bytes
Raw data transmitted
Paillier HE + edge inference: raw sensor values never leave wearable. Invariant I1 enforced at hardware level.
5-tier
Resilience stack
5G → WiFi → LoRaWAN → Satellite → Store-and-forward. Never fully blind. HE participation effect: +1.8 pp indirect accuracy gain.

Oslo Safe Schools Declaration — alignment

PoA blockchain implements SSD data collection mandate for attack reporting. Edge autonomy implements operational continuity. Privacy-by-Design respects civilian character. 100% audit provides legally admissible incident records for the Nigeria Safe School Initiative.

// performance radar — CEICS vs cloud

// ablation — accuracy contribution per component

Large-scale stress test — 1000 devices · 20 schools · 3 locations · 30% failure

Average alert latency
Alert delivery success
Gateway CPU load
Packet loss rate
Satellite fallback usage
Privacy enforcement
100%

// stress latency vs cloud

// delivery success vs node count