Business Problem & Solution

A property's perimeter is covered by cameras that almost never show anything worth acting on. A human watching a wall of mostly-static feeds habituates within minutes; running a model on every frame of every camera is expensive and still drowns the operator in benign motion — staff crossing the loading bay, a fox on the north fence, a guest in the lobby. The signal that matters — a person dwelling in a restricted zone, a vehicle breaching the perimeter — is rare, and rarity is precisely what continuous monitoring is worst at surfacing.

App 09 is the control room for that problem. It samples each camera's frames, cheaply skips the static ones, runs detection only on frames that changed, filters out the classes that are expected for a zone, and confirms a candidate only after it dwells across a window of frames — so a single-frame false positive never raises an alert. What reaches the operator's queue is a small number of confirmed, severity-rated events, each one explainable down to the deciding frame.

It proves the cost-and-noise discipline: a change-gate short-circuits near-static frames at $0 before any metered detection, and a dwell window turns flickers into confirmed tracks. The result is that perception spend scales with activity, not with camera count, and the alert queue stays legible. A confirmed event is raised to the queue and emitted as `security_event.raised` on the shared bus — App 09 is the producer of C5's security spine.

It simplifies the vision itself. The frames are synthetic and ICONOGRAPHIC — there is no real footage, no faces, no people; objects are abstract markers on a top-down field (privacy by framing). The `perceive` stage is a SIMULATED metered detector reading the frame's ground-truth objects. The sampling, change-gate, class filter, zone test, dwell confirm, rules and alerting are all real pipeline logic; only the model inference is stood in for.

Synthetic data only — three invented properties, each with cameras on its perimeter, restricted and lobby zones, and a short deterministic frame loop per camera (~14 frames, most of them static). No real footage, faces, people, brands or PII — ever. Detection classes (person · vehicle · staff · animal), confidences, dwell seconds and per-frame costs are representative and labelled. The cost meter, mode and per-stage trace are confined to the dark inspector.