Parlant / Output side
A typed replay adapter turns a normalized Parlant activation into an existing coaching advice record. It does not select guidelines or make safety decisions. The real coaching pipeline keeps arbitration ownership.
Showcase 01 · simultaneous conflict
One normalized security activation at priority 10 and one follow-up at priority 1 arrive 100 milliseconds apart in the same session.
The real coaching pipeline retained only the priority-10 safety instruction for that decision moment.
Without the priority-10 activation, the priority-1 follow-up is the remaining candidate for that same moment.
Showcase 02 · later instruction
The same priority-10 security activation arrives at 1,000 milliseconds; the priority-1 follow-up arrives at 2,001 milliseconds.
The pipeline emitted the safety instruction and then the follow-up: 1,001 milliseconds is outside its one-second simultaneity window, so this is a new decision moment.
Without the later activation, the replay emits only the first safety instruction. Nothing invents a second message.
These are deterministic replay tests over a normalized boundary. They prove the adapter/pipeline path and its defined arbitration behaviour, not that a live Parlant deployment is installed in a production coaching session.