raxIT Labs evaluated eighteen decision models against Bedrock Guardrails over 9,975 checks across six suites covering eight use cases. Scoring uses macro-averaged balanced accuracy under a uniform default rule where a probability of 0.5 or greater blocks input. Models did not undergo per-task threshold tuning, and the text-only benchmark used Jev's native question format. Leaders varied significantly by category, meaning performance in one area does not guarantee general replacement capability.
How you can use it
Use the published checks and per-row results to see which kinds of mistakes matter for your app, rather than choosing from the overall ranking alone. A developer can rerun providers with the same questions and blocking rule. Compare false alarms as well as missed hazards; text-only results do not establish image safety.
