Every request carries three hidden risk scores in [0,1]: P (policy violation), M (moderation / toxicity) and Q (PII presence). Each gate compares its own score to a threshold set by the sliders — a higher strictness slider lowers the threshold, so more requests trip the gate. A tripped gate either blocks the request outright or transforms it (masks PII, softens the score) and lets it continue, flagged. The final enforcement gate recomputes a combined risk score: R = 0.4·P + 0.4·M + 0.2·Q and blocks anything still over the working threshold.

• Policy strictness — threshold for the policy-engine gate (τ_p)
• Moderation sensitivity — threshold for the toxicity classifier (τ_m)
• Traffic rate — how many requests/second enter the pipeline
• PII mode — mask detected PII in place vs. block the request
• Enforcement mode — transform-and-continue vs. hard block on violation

In production this maps to a policy engine + moderation classifier + PII scrubber sitting between the user and the model, logging every decision for audit and tuning.