Under the hood: LightGBM, ~80 features and sub-100ms scoring

A single check does a lot of work in a very short time.

1. The widget or API bundles the form fields and (with the widget) behavioural signals into JSON.
2. Veritus fans out concurrent enrichment lookups across the seven signals. Most are answered from a Redis cache with a hit rate above 90% in steady state; misses go to the underlying provider with short TTLs. Typical enrichment completes in under 60ms.
3. A LightGBM gradient-boosted classifier scores roughly 80 features (after one-hot encoding) in about 2ms, producing a 0–1 probability.
4. The probability maps to a verdict, SHAP-style attribution produces the reason codes, and the full response returns.

End-to-end latency runs around p50 73ms, p99 under 150ms. Slow third-party vendors time out without holding up the check.