Research

Benchmarks · Aug 2026

Measured, not estimated.

Every number below comes from a real test run against the live system — a k6 load test against the production ingestion endpoint, and validation metrics from the actual trained models. No projections.

Ingestion load test

k6, ramping 1,000 → 2,500 → 5,000 concurrent virtual users against /v1/ingest.

11ms
p95 latency
5,000
Peak concurrent users
1K → 2.5K → 5K
Ramp stages
0%
Error rate

Model validation scores

Trained on a dataset of ~2,724,000 labeled sessions. See the whitepaper for the full methodology.

M3 — XGBoost bot classifier

Validation AUC

0.9993

M2 — Siamese LSTM (keyboard identity)

AUROC

0.9673

M4 — Autoencoder (mouse + scroll)

AUROC

0.9993

M4 — Autoencoder (touch + motion)

AUROC

0.9976

Training data composition

~1,974,000 human sessions (keyboard-only, mouse/touch/scroll modalities at zero — the empirically correct representation for single-page sessions) and ~750,000 bot sessions, spanning all six behavioral modalities.