AI Guardrail Pipeline 2D: Signal-Detection Simulator

Filter Thresholds

Input filter strictness 0.55

Output filter strictness 0.55

Adversarial evasion skill 0.20

Traffic rate (prompts/s) 2.0

Detector noise σ 0.12

Live Metrics

Incidents (leaked)

Overblocked benign

Block rate

Latency overhead

40 ms

Precision

—

Recall

—

How it works

Every incoming prompt carries a hidden ground-truth risk r ∈ [0,1] (harmful prompts sample r ~ Uniform(0.55, 1), benign ones r ~ Uniform(0, 0.45) — real distributions overlap, which is why no filter is perfect). Each gate never sees r directly, only a noisy observation:

score = r + N(0, σ) − evasionSkill × 0.35 × isHarmful
decision: BLOCK if score > threshold, else PASS

This is classic signal-detection theory: raising a gate's threshold catches more harmful prompts (higher recall) but also blocks more benign ones (lower precision — "overblocking"). A skilled adversary subtracts from their own score to slip under the threshold ("evasion"), which is why real guardrail stacks use two independent layers rather than one perfect gate.

The lower panel plots every evaluated score as a live, decaying histogram — green for benign, red for harmful — with two draggable vertical handles marking the input and output thresholds directly on the distributions they act on. Drag a handle (or use its slider) and watch the overlap area it now covers change what gets blocked.

Precision = TP / (TP + FP)
Recall    = TP / (TP + FN)
Incident  = a harmful prompt that passed every stage (false negative at the end)

Input filter — checks the raw prompt before it reaches the model core.
LLM core — pass-through marker; no blocking happens here, it represents generation.
Output filter — a second, independent check on the generated response.
Monitor — logs everything that reaches this point but never blocks (observability, not enforcement).
Latency overhead — every inspection costs compute; stricter thresholds run deeper checks, so overhead scales with both thresholds.

Drag horizontally on the top lane view to pan the camera along the pipeline, or scroll to zoom — both panels always drive the exact same underlying model, just two different lenses onto it.

Benign prompt Harmful prompt Blocked at a gate

drag lane view to pan · scroll to zoom
drag histogram handles to set thresholds