Start with one consequential path.
RuleOak works best when the action matters, the boundary is under your control and you can define what evidence should exist after execution.
Filesystem mutation through MCP
Wrap a supported MCP route, authorize writes/deletes and use reviewed filesystem postconditions where available.
Deployment / infrastructure change
Bind approval to the exact target/action and verify the actual deployed state—not merely command exit status.
Outbound API/webhook action
Prevent blind duplicate retry after a lost response by checking a stable external identifier/postcondition where the API supports it.
Guarded Node.js tool registry
Keep authorization, lease consumption, execution, EffectReceipt and evidence in one controlled call path.
AI-produced source candidate
Review an external candidate against declared intent and baseline-owned qualification without making the producing model its own verifier.
Reliability qualification
Use CrashLab to exercise response loss, duplicate retry, revocation, expiry, stale state and concurrency.
Readiness questions
- Can you prove the consequential action actually crosses RuleOak?
- Can deny or approval-required stop execution?
- Is one-use authority consumed before the side effect?
- Is there a reviewed external postcondition for the effect?
- What happens when that postcondition is unavailable or ambiguous?