AI Guardrail Pipeline 2D: Signal-Detection Simulator

Production LLM systems don't rely on a single perfect filter — they chain independent checks: an input filter before the model, the model itself, an output filter after it, and a monitor that logs everything without blocking. This 2D simulator renders that pipeline as a top-down lane view of prompts streaming through detection gates, paired with a live histogram of the noisy risk scores each gate actually sees. Each prompt carries a hidden true risk score; each gate observes risk plus detector noise minus any evasion penalty, and blocks anything above its threshold — drag the histogram's own threshold handles, or the sliders, to move it. Push the input and output thresholds up and incidents (harmful prompts that slip through every gate) drop — but so does throughput, as more benign prompts get caught in the crossfire (overblocking). Raise the adversarial evasion slider or fire a red-team burst to see how a skilled attacker erodes a single gate's effectiveness, and why real guardrail stacks stack multiple independent layers instead of trusting one.