The escalation ladder

   Patrol  →  Investigate  →  Pursue
   (idle)   (suspicious)    (engaged)
      ↑           ↑              |
      |           ↓              ↓
      ←————————————  cooldowns  ——

Each tier has higher priority than the one below. De-escalation isn't a hard transition — it's the higher-tier goal's score dropping below the lower-tier's after enough time / distance has elapsed. The goal commit floor (Family 5) makes de-escalation deliberate rather than instantaneous.

Anatomy

Sensor fusion lives in three sensor components attached to the archetype: a sight cone, a hearing sphere, and a damage listener. Each writes its own fact. The latched bools (PlayerVisible, HeardSomething) are what the goals actually score against.

Why this example is here

Three production lessons:

1. Sensor fusion via separate facts beats trying to compute a single "threat level" scalar. Each sensor is independent; the planner combines them via goal scoring.
2. Time-since-contact as a derived scalar gives natural de-escalation behavior. After 15 seconds without re-confirmation, the Investigate goal's score drops below Patrol's and the guard transitions out — without hand-authoring "if 15 seconds elapsed, stop investigating."
3. 8-second commit floor on Pursue matters. Without it, the guard would flip back to Investigate every time the player ducks behind cover (PlayerVisible→false), even though it has every reason to keep pursuing.

At no point should the guard rapidly flicker between Pursue and Investigate when the player is intermittently visible.