See a change before you make it.
A visibility policy is only as good as your confidence in it. Because every decision runs through one engine, you can ask it why, rehearse a change, and replay the past, all without touching what is live.
Stop guessing what a policy change will do.
Access rules are easy to get subtly wrong, and the cost of getting them wrong is an agent seeing something it should not. The sandbox turns policy from something you edit and hope into something you interrogate, preview, and check. Every tool here reads and evaluates through the same engine that enforces, so the preview can't drift from the rules that enforce.
Every allow and deny says why.
Each decision carries a plain-language reason and names the exact rule that produced it, so you can see why an item was shown, masked, or withheld, without reading code.
Preview a change without applying it.
Toggle a hypothetical clearance, ceiling, or role and watch the effect at once. The same engine evaluates the hypothetical, so the preview cannot drift from what enforcement would actually do, and nothing about the live policy is touched until you decide.
reviewer → proposal : blind
reviewer → proposal : blind → redacted tighter
Test a change against real traffic first.
Run a proposed change against a body of recorded requests and read the aggregate effect, so you know the blast radius before you flip anything on.
8 of 19 recorded requests would change (42%)
proposalPolicy has a history, and you can travel it.
Rollback with a diff
Every policy is a version. Preview reverting to a prior one and see exactly what it would re-open or close, before you commit to it. This is forward-looking: reopening a clearance changes future decisions; it does not un-see what an agent already read or remove it from memory.
Time-travel replay
Re-decide past requests as of an earlier policy version, to see which of them would have gone differently than they do now. An audit tool that answers "what would this have allowed back then."
Catch what leaks around the edges.
Lineage shows you how information actually moved between agents. If something an agent was excluded from still reached it indirectly, through another agent's summary or something derived from it, lineage surfaces that path and flags it for review, so indirect influence is visible instead of silent.
One engine, asked instead of edited.
The sandbox is not a second copy of the rules that could drift from the real ones. Each tool passes a hypothetical to the same policy engine that enforces on the path, reads back the decision, and shows it. Previews are input only: they never mutate the live policy. That is why what you rehearse in the sandbox matches what the same policy enforces in production.