Empirical Quality & Speed Audit
100,000 Ground-Truth Email Benchmark
Most email verification vendors make vague claims without publishing raw evaluation datasets. Here are our exact empirical results across 100,000 ground-truth emails and 9 threat categories.
99.85%
Overall Accuracy
99,847 / 100,000 passed
99.69%
Threat Recall
49,847 / 50,000 bad caught
< 0.3ms
Single Request Latency
Pure compiled Rust
3,203/s
Batch Throughput
31.2s for 100k emails
Audit Verification
N = 100,000 Ground-Truth SamplesOfficial Confusion Matrix
| Ground Truth | Predicted BAD (Deny) | Predicted GOOD (Allow) | Accuracy Rate |
|---|---|---|---|
| Actual BAD Actors (50,000) | 49,847 True Positives (Blocked) | 153 False Negatives (Missed) | 99.69% Recall |
| Actual GOOD Users (50,000) | 0 False Positives (Blocked) | 50,000 True Negatives (Allowed) | 100.00% Specificity |
Our Production Philosophy: While 0 legitimate human accounts in our 50,000 benchmark set triggered a hard
DENY, real-world traffic always includes ambiguous edge cases. EmailScore routes uncertain or borderline addresses to CHALLENGE (score 40–69), allowing you to prompt for soft verification (OTP / email link) without ever blind-blocking paying customers.Granular Evaluation
Accuracy Breakdown Across 9 Threat Categories
Every subcategory tested against known threat models and real-world corporate corpora.
| Threat / Account Sub-Category | Sample Size | Expected Decision | Correct Detections | Accuracy |
|---|---|---|---|---|
| Corporate Professional Aliases | 15,000 | ALLOW | 15,000 | 100.00% |
| Human Multi-Token Webmail (Gmail/Yahoo/Outlook) | 25,000 | ALLOW | 25,000 | 100.00% |
| Human Natural Suffixes (e.g. alex1994) | 10,000 | ALLOW | 10,000 | 100.00% |
| Disposable Burner Domains | 25,000 | DENY | 25,000 | 100.00% |
| Dead Domains (NXDOMAIN) | 1,250 | DENY | 1,250 | 100.00% |
| Parked Domains (Sedo, Bodis, Dan) | 1,250 | DENY | 1,250 | 100.00% |
| Trial Churn & Subaddress Salt Abuse | 5,000 | DENY | 5,000 | 100.00% |
| Burner DGA Domains (Algorithmically Generated) | 5,000 | DENY | 4,958 | 99.16% |
| Bot PRNG Hashes (e.g. x89qj2k98a1@gmail.com) | 12,500 | DENY | 12,389 | 99.11% |
Market Comparison
EmailScore vs Legacy Email Verification APIs
Why traditional bulk newsletter validators fail on the signup hot path.
| Evaluation Dimension | EmailScore (emailscore.io) | ZeroBounce | NeverBounce | AbstractAPI |
|---|---|---|---|---|
| Primary Architectural Focus | Real-Time Signup Defense & Anti-Bot | Bulk Newsletter Cleaning | Bulk List Scrubbing | Data Enrichment |
| Response Latency (Hot Path) | 0.3 ms (In-memory Rust) | 1,500 – 4,000 ms | 1,200 – 3,500 ms | 800 – 2,000 ms |
| Catches Webmail Bots (`*@gmail.com`) | ✅ Yes (99.11% Recall) | ❌ No (Whitelists @gmail.com) | ❌ No | ❌ No |
| Catches Zero-Day Burner Domains | ✅ Yes (In-Memory DGA + Entropy) | ❌ No (Relies on static lists) | ❌ No | ❌ No |
| Data Retention & Storage Policy | 🛡️ 100% Stateless (Never Stored) | ⚠️ Stored & Logged | ⚠️ Stored & Logged | ⚠️ Stored & Logged |
| Batch Throughput | 3,200+ emails / second | ~20–50 emails/sec | ~30–60 emails/sec | ~25 emails/sec |
| Risk of Server IP Blacklisting | Zero (No SMTP RCPT TO probes) | High (Aggressive SMTP pings) | High | Moderate |
Open Verification
Reproduce the Benchmark Locally
We make our test harness transparent. You can clone the repository, compile the release engine, and run the 100,000-sample empirical verification on your own machine.
git clone https://github.com/emailscore/emailscore.git
cd emailscore
./scripts/run_quality_benchmark.sh
cd emailscore
./scripts/run_quality_benchmark.sh