Empirical accuracy.
Rigorous confusion metrics.
Statistical precision verified against an independent, non-circular test corpus evaluated across the production Sentinel inspection engine and hydrated Global Threat Grid. Every metric reports exact True/False Positives and Negatives with 95% Wilson score confidence intervals and base-rate prevalence modeling.
Threat Cohort Sensitivity & Recall (N = 172,500)
Evaluated against ground-truth adversary ingress endpoints. Sensitivity / Recall reflects the probability of correctly identifying a genuine proxy, VPN, or exit relay (TP / (TP + FN)).
| Target Cohort | Sample (N) | True Pos (TP) | False Neg (FN) | Recall (TPR) | 95% Wilson CI | Independent Ground Truth |
|---|---|---|---|---|---|---|
|
Commercial VPN Gateways
Consumer & Enterprise WireGuard/OpenVPN
|
50,000 | 49,937 | 63 | 99.87% | [99.84%, 99.90%] | Provider config harvesting & active canary ingress tunnels |
|
Datacenter VPS & Cloud Transit
AWS, GCP, OVH, Hetzner, Linode, DigitalOcean
|
75,000 | 74,982 | 18 | 99.98% | [99.96%, 99.98%] | Carrier Network Classification & Regional Internet Registry WHOIS |
|
Tor Consensus Exit Relays
Sentinel Tor Relay Intelligence
|
2,500 | 2,500 | 0 | 100.00% | [99.85%, 100.00%] | Cryptographic directory consensus published out-of-band |
|
Residential Proxies & Covert Tunnels
Dynamic peer exit nodes, SDK bot relays
|
45,000 | 44,262 | 738 | 98.36% | [98.24%, 98.47%] | Sentinel Global Threat Grid & Dynamic Tunnel Anomaly Detection |
| Threat Subtotal | 172,500 | 171,681 | 819 | 99.525% | [99.492%, 99.557%] | Overall Target Threat Recall |
Clean Control Specificity & False Positive Rate (N = 77,500)
Evaluated against clean consumer residential connections. The False Positive Rate (FPR) measures how often legitimate players are mistakenly flagged (FP / (FP + TN)), disentangled from class precision.
| Control Cohort | Sample (N) | True Neg (TN) | False Pos (FP) | Specificity | FPR | 95% Wilson CI (FPR) | Ground-Truth Verification |
|---|---|---|---|---|---|---|---|
|
Clean Residential Consumer Broadband
FTTH / DOCSIS / Fixed Wireless Tier-1 ISPs
|
55,000 | 54,975 | 25 | 99.955% | 0.045% | [0.031%, 0.067%] | 90-day device-attested human gaming sessions |
|
Clean Mobile Cellular (CGNAT Pools)
4G LTE / 5G SA Shared Carrier NAT IP Pools
|
22,500 | 22,468 | 32 | 99.858% | 0.142% | [0.101%, 0.201%] | Subscriber Carrier Verification & Mobile Network Codes |
| Control Subtotal | 77,500 | 77,443 | 57 | 99.926% | 0.074% | [0.057%, 0.095%] | Aggregate False Positive Rate across Controls |
Bayes' Theorem & Real-World Precision
In our 250,000 endpoint test corpus, proxy traffic prevalence is 18.0%. However, a typical Minecraft or game server observes only 0.5% to 2.0% proxy connection attempts. Under Bayes' theorem, Positive Predictive Value (PPV) depends heavily on real-world prevalence:
PPV = (TPR × p) / [ (TPR × p) + FPR × (1 - p) ]
| Inbound Proxy Prevalence | Network Scenario | Overall Threat PPV | Shared Mobile CGNAT PPV | Recommended Defense Posture |
|---|---|---|---|---|
| 0.5% proxy traffic | Whitelisted / niche community server | 87.18% | 77.66% | Graduated friction (CAPTCHA / 2FA verification) |
| 1.0% proxy traffic | Typical public Minecraft server | 93.18% | 87.48% | Hard kick VPN/VPS; challenge residential tunnels |
| 2.0% proxy traffic | Competitive PvP / Factions / Anarchy network | 96.51% | 93.38% | Enforce risk threshold ≥ 80 |
| 5.0% proxy traffic | Tournament / high-stakes ranked match | 98.62% | 97.33% | Strict enforcement with staff bypass role |
| 18.0% proxy traffic | Our empirical benchmark corpus | 99.66% | 99.35% | Raw test set baseline |
Cellular carriers assign thousands of mobile subscribers behind a single egress IP via Carrier-Grade NAT (CGNAT). If a malicious actor on that cell tower routes proxy traffic, flagging that carrier IP can inadvertently impact real cellular players. This is why Sentinel provides granular confidence and reason_codes: developers can hard-kick datacenter proxies while issuing non-destructive authentication challenges (such as web authentication or Discord linking) for residential tunnel signals.
Independent Ground-Truth Collection
Testing an engine against its own internal database yields circular 100% results. To guarantee objective accuracy, our 250,000 endpoint validation corpus was compiled out-of-band and evaluated across the live Sentinel inspection engine:
Residential Proxies
45,000 residential proxy targets combining active nodes from the Sentinel Global Threat Grid with dynamic tunnel sessions exhibiting transport and temporal anomalies.
Commercial VPNs
50,000 VPN exit endpoints compiled from WireGuard, OpenVPN, and IKEv2 gateway configurations across major commercial VPN providers and unlisted regional relays.
Tor Exit Relays
2,500 active exit nodes verified through Sentinel Tor Relay Intelligence and Anycast edge country classification.
Clean Control Set
77,500 clean IPs curated across tier-1 consumer fiber/cable ISPs (Comcast, AT&T, Verizon, Deutsche Telekom, BT, Orange, Vodafone) and 4G/5G CGNAT mobile carriers.
Stateless Edge Execution vs. Background Threat Grid
A common technical question: how does Sentinel achieve "stateless, sub-millisecond edge evaluation" while maintaining "temporal correlation and risk decay"? The answer lies in strict pipeline separation:
Zero-Log Volatile Memory Architecture
When an API request hits our Anycast edge, evaluation executes within a V8 isolate that accesses a compiled, immutable binary threat store residing entirely in RAM.
- ✓ Zero database queries during request evaluation
- ✓ Zero disk writes or session persistence
- ✓ In-RAM binary search & prefix matching in 0.0068ms mean compute
Sentinel Global Threat Grid Pipeline
Temporal correlation, velocity anomaly detection, and risk decay execute asynchronously in our background Threat Grid aggregation cluster.
- ✓ Dynamic tunnel churn tracked across time series graphs
- ✓ Inactive proxy IPs decay exponentially over 12–72 hours
- ✓ Compiled immutable binary snapshot broadcast to edge isolates
Reproduce These Metrics Locally
We provide an open validation script that probes our live Anycast edge API in real time, validates each cohort, computes the confusion matrix, and prints Wilson score 95% confidence intervals: