You learn more from one session run in front of real participants than from a month of authoring. These are the notes I took from the back of the room, lightly cleaned up, on the first day insigz carried a full exercise.
Briefing
The room settled faster than I expected. The shared situation display — the live map of the world-state — did something a slide deck never does: it gave everyone a single thing to point at. Within minutes people were arguing about the same map instead of talking past each other about different mental models. That alone justified the build.
First turn
The first cell submitted an action and waited for the world to respond. The Adjudicator drafted a consequence with its reasoning attached, faculty edited one sentence and approved it, and it committed to the timeline. The edit mattered: the faculty member softened a claim the agent had stated too confidently. That is the human-in-loop loop working exactly as designed — the agent drafts, the human commits, the human's judgment is the one that ships.
The platform earned its keep the moment the room started arguing about the same map.
Where it got in the way
Two friction points. First, the proposal inbox buried the grounding one click too deep; faculty wanted the cited observations visible before deciding, not after expanding. Second, an inject fired a beat late because the scheduler and the turn clock weren't tightly coupled, and a late inject is worse than no inject — it breaks the room's sense of cause and effect.
The debrief
The structured debrief, with the timeline replayed and the key decisions surfaced, was where the learning crystallized. Participants could see not just what they decided but when they had the information to decide it. The after-action report landed afterward as a signed PDF — the artifact that leaves the room and becomes the next class's reading.
What to fix next
- Surface grounding in the proposal inbox by default, not behind an expand.
- Couple the inject scheduler to the turn clock so timing is exact.
- Give faculty a one-key "narrate" mode to address the room without leaving the umpire view.
None of these are architecture problems. They are the small, specific frictions you only find by watching a real room use the thing. Which is the entire argument for running one session before building the next feature.