Drew a firm boundary between demo convenience and deployable defaults. Every prototype accumulates shortcuts. The question is whether you acknowledge them or pretend they do not exist. We are acknowledging them and putting them behind an explicit sandbox flag.
Started splitting Pilot behavior from Alpha behavior. Pilot is the learning environment where teams can see how the system works, test policies against realistic scenarios, and build intuition for the evaluation model. Alpha is where real actions hit real systems. The gap between those two is bigger than it looks.
Added sandbox conventions that make shortcuts explicit and visible. If something is a sandbox-only behavior, it says so. Loudly. No silent degradation where a demo convenience quietly becomes a production default because someone forgot to flip a flag.
Built deterministic stub execution for connectors so that Pilot mode never makes real outbound calls. The stubs return realistic responses but nothing actually happens. This means teams can test their entire policy configuration against real-looking traffic without worrying about side effects.
Pilot is about learning safely. Not proving production readiness.