Fault Injection & Test Gates Lab
What happens when a protective gate is absent, and what does a test receipt really establish?
Browser-local learning exercise. All scenarios, costs and receipts are synthetic. No real model, API or service is being evaluated.
1. Inject a synthetic fault
Choose a known failure fixture. Each scenario represents a simplified model of how a particular control would behave.
2. Toggle the five test gates
Turn off a gate to let its known fixture pass. Select a stage to view the receipt.
3. Inspect the explanation and receipt
Model assumption
Control's intended action
Verification receipt - synthetic
4. Try a defense-in-depth experiment
Choose Weak RAG evidence. Disable Gate 2 while keeping Gate 5 on: the unsupported claim passes retrieval but is stopped by the later grounding check in this model. Disable Gate 5 too: the model reports an unhandled risk.
No LLM, guardrail, RAG service, network or tool is actually executed. A real classifier might miss an injection; a simulated interception is never evidence of deployed protection. Latency and token usage are deliberately not fabricated.