Each request passes through a stack of independent guardrail layers — allow/deny lists, content filters, and PII / runtime monitors. Every active layer gets one chance to catch a malicious request, so the odds of catching it at least once rise fast with more layers.
P(caught) = 1 − (1 − s)^n
P(overblocked) = 1 − (1 − s·k)^n , k ≈ 0.12
- Request rate — how many requests per second enter the pipeline.
- Malicious share — the fraction of incoming requests that are actually adversarial.
- Sensitivity (s) — per-layer detection strength; higher catches more threats but also blocks more legitimate traffic (recall vs. precision).
- Guardrail layers (n) — toggle layers off to see coverage gaps open up immediately.
Red spheres are malicious requests, blue-green are legitimate. A flash at a gate means it was caught there; a request that clears every active layer flies on to the model core and lights it up.