Skip to content

Enforcement Simulator

5. Sandbox (Runtime Enforcement Simulator)

The Sandbox is the Runtime Enforcement Simulator — an interactive non-production environment for testing governance logic, modifying inputs, simulating edge cases, and exporting audit-ready JSON before deploying policies to production agents.

5.1 Simulation Input

Users configure a simulation run by specifying:

  • Request Text — The prompt or user message to simulate
  • Response Text — The AI-generated response to evaluate
  • Guard / Domain — Which guard kernel to evaluate against
  • L1 Score Override — Optionally set a specific L1 divergence score to test boundary conditions
  • L2 Verdict — Pre-set the L2 verdict for the simulation
  • Watchpoints — Named behavioral flags to trigger during evaluation
  • Tool Calls — Simulate tool invocations by the agent
  • SSC Severity — Structural Safety Check severity level

5.2 Simulation Results

After running, the Sandbox returns:

  • Enforcement Decision — The action the system would take (allow, block, rewrite, route to review)
  • L1 Divergence Score — The semantic shift score for the simulated interaction
  • L2 Confidence — The L2 classifier's confidence in its verdict
  • L3 Triggered — Whether the L3 LLM evaluator would be activated
  • Blocked Tools — Any tools that would be blocked by the guard
  • Applied Guards — The guards that would be applied to this interaction
  • Safe Degrade — Whether a safe degraded response would be substituted

5.3 Adversarial Test Cases

The Sandbox supports running predefined adversarial prompt libraries against a guard to measure how robustly it handles known attack patterns (prompt injection, policy circumvention, data exfiltration probes).