Test Set Accuracy
160/160 Validated
100.00%
Macro F1
EXP-03 BiGRU: Zero test misclassifications on held-out test suite.
Active Parameters
-79.2% Slashed
6,544
Parameters
Streamlined architecture slashed from 31,516 in Modified BiLSTM.
Runtime Footprint
INT8 Dynamic
15.1 KB
Disk / RAM
97.8% footprint reduction for extreme edge MCU & IoT gateways.
Inference Speed
3.9x Faster
0.36 ms
per inference
Zero cloud latency; runs 100% locally in browser / client runtime.
1. Interactive Session Studio
Quick Test Scenarios:
Clickable Event Palette (Click to append):
Device:
Live Reactive
CRITICAL RISK
99.8% Match
Suspicious Retry Loop
Abnormal cluster of RETRY tokens detected following assessment access. Indicates credential brute-forcing, script automation, or exam tampering.
Automated Campus Response:
POLICY-SEC-04
Force Step-Up MFA & Rate-Limit IP (5 req/min). Log incident to campus SOC.
Regime Probability Distribution:
🚨 retry_loop (Anomalous)
0.0%
✅ smooth (Standard Flow)
0.0%
⏳ idle_abandon (Terminal Risk)
0.0%
🔍 explore_heavy (Research)
0.0%
Trigger Token Heatmap:
<PAD> masked to 0% weight via Attention
Challenge Card C4: Empirical Hardware & Compute Trade-Off Table
| Model ID | Architecture | Parameters | Weight Size | CPU Latency | Held-Out Test F1 | Target Deployment |
|---|---|---|---|---|---|---|
| EXP-01 Baseline | 1-Layer LSTM (Naive pooling) | 6,404 | 30.1 KB (FP32) | 0.32 ms | 98.12% (3 padding errors) | Unoptimized Reference |
| EXP-02 Modified | 2-Layer BiLSTM + Masked Attention | 31,516 | 679.2 KB (FP32) | 1.42 ms | 100.00% (160/160 correct) | Centralized Cloud Sentinel |
| EXP-03 Challenge (Active) | 1-Layer BiGRU + Streamlined Attention | 6,544 (-79.2%) | 15.1 KB (INT8 Quantized) | 0.36 ms (3.9x faster) | 100.00% (Zero loss) | Campus Lab Edge MCU / Router |