Try it out

Try it out is a simulation. No live agent action is executed. Use it to answer "what happens if I make this change?" before you make it.

It has four modes, all available in the console under Try it out and through the API.

Mode	The question it answers	Calls
Request simulation	What would Bairav decide for this agent and action? What happens if I change a policy, change an agent's access, or tighten a guardrail?	POST /api/policies/simulate, POST /api/simulations/rule-change, /access-change, /guardrails
Blast radius	What could a compromised agent reach?	GET /api/simulations/blast-radius
Exposure scoring	How bad would that be?	GET /api/simulations/exposure, /exposure/fleet
Adversarial testing	Does my configuration hold against known attacks?	POST /api/simulations/adversarial, /adversarial/fleet, /adversarial/start

Request simulation

POST /api/policies/simulate returns the decision Bairav would make and the policy that would make it. It is rate limited per tenant and user, and the AI-judge call fails closed: if the judge cannot be reached the result says so rather than guessing.

Related calls:

• GET /api/policies/simulate/examples returns real material for starting a scenario.
• GET/POST /api/policies/scenarios and POST /api/policies/scenarios/run keep and re-run saved test cases.
• POST /api/policies/replay re-checks past gateway decisions against today's policies.
• GET /api/simulations/publish-check checks a policy before it is published.

Long runs

Adversarial runs across the whole fleet run as a job. Start one with POST /api/simulations/adversarial/start, poll GET /api/simulations/jobs/{run_id}, cancel with POST /api/simulations/jobs/{run_id}/cancel, and save the result with POST /api/simulations/jobs/{run_id}/record.

Save results

POST /api/simulations/record (or /runs) records a run so it can be reviewed later, and GET /api/simulations lists them.

Who can run it

Authoring simulations need the admin or reviewer role. Posture simulations (blast radius, exposure) are available to a wider set of roles. These calls need a signed-in user session; machine tokens are not accepted.

Honest limits

• A simulation shows the decision logic and its evidence. It is not a promise about the next real request, because live data can change between the two.
• Where a result depends on the AI judge and the judge is unavailable, the response says it was unable to evaluate rather than returning a pass.