On the left, six software agents depend on each other and fail. On the right, a simulated person meets cues that raise distress. The same Watchdog chooses each side's working mode every tick, and the Spine and the safety layer can overrule it.
These are replays of real runs of the tested code (rsxv: World, AgentEnvV2, PersonEnvV2), recorded tick by tick. Both sides are stated models, not people and not production systems. Choose an intervention and a run, press play, and click the timeline to see why any decision was made.
Each circle is an agent; its ring shows how wrong its output is. Arrows point from an agent to the agents it depends on, so one agent's fault spreads downstream.
Distress, arousal and agency of the simulated person, with the safety mode in force.
Gold line: agent system health. Coral line: the person's distress. Coral dots: cues reaching the person. Gold marks: the safety layer overruling the Watchdog. The shaded band is the intervention. Click anywhere to jump to that tick.
The decision chain for the person this tick, then for the agents.
Every switch, cool-down, promotion and sleep is a typed edge with a reason; any edge can be rolled back from a snapshot.
The Watchdog estimates, from what it has seen so far, which mode fits each context, and switches through a two-phase handover. It also micro-randomises some ticks to keep learning. On the person side, a safety mode such as STABILISE blocks deep work whatever the Watchdog wants, and the run falls back to stabilising. Each block shows as a small gold mark on the timeline and as a step in the decision chain for that tick.
Results of every registered test, passes and failures, are on the results page. The confirmed result most relevant here is T-AP: a policy learned on agents, certified and gated, helps the simulated person.