The spec that decides whether a theft detection system works is the raw-to-delivered alert ratio, not the accuracy number.

Every camera on a property generates raw events all day: a person walks past a lobby, a car crosses a lot, a dog runs across a lawn. A theft detection system that forwards all of those to a human is a system that gets muted. The filter ratio is the number of raw camera events consumed for every delivered alert. On a 16-camera multifamily property, a typical day produces 200 to 300 raw person detections. A system keeping its alert channel usable delivers 3 to 8 alerts. That is roughly a 50-to-1 compression ratio, and it is achieved by stacking three scoped filters on top of person detection: a polygon zone, a dwell threshold, and a time window.

Because accuracy on its own is not a deployment metric. A model that is 99 percent accurate but fires 150 times a day at a 16-camera property will have its WhatsApp thread muted by the property manager before the week is out. Once the channel is muted, the true positives go unseen too. The system is now a recording, not a detection product. In the field, the question that actually determines whether alerts continue to be read is: how many events does the staff see per shift? Keep that number in the single digits per day and the channel stays active. Push it into the dozens and the mute happens every time.

Three filter layers running after person detection. First, the polygon zone: on every camera tile at install, operators draw the boundaries that matter, mailroom door, transformer pad, HVAC cage, dock apron, fire lane. A person outside every drawn polygon is not an event. Second, the dwell threshold: the person must stay inside the polygon longer than the threshold, typically 4 seconds or more. A resident walking past a transformer in 2 seconds never crosses it. Third, the time window: each zone has an armed schedule, and an alert only fires when the current time is inside it. A maintenance crew on the pad at 10 a.m. during a scheduled window is not an event. The three filters combine multiplicatively, which is why the compression ratio compounds so aggressively.

A DVR motion alarm triggers on pixel change. A raccoon, a tree shadow at sunset, a plastic bag blowing across the lot, and an actual person are all pixel changes. Most DVR motion alarms wake staff dozens of times per night, which is why most of them are disabled within a month. Cyrano runs object-aware person detection first, which removes the entire animal, weather, and lighting class of false positives. Then the zone, dwell, and time filters run on top of that. The result is that the same 16-camera property that generated 40 to 60 motion alarms a night generates 3 to 8 zone-verified alerts a day, every one of which is a person inside a polygon longer than the dwell threshold during an armed window.

Each delivered event carries a tile thumbnail (cropped from the DVR multiview), the zone label that was crossed, the dwell seconds counted, a timestamp, the camera name, the DVR's layout_id (e.g. 4x4-std or 5x5-std, used to apply the overlay mask at inference time), and the end-to-end latency in milliseconds from frame capture to message send. The event class for a pre-action pattern is pre_action_zone_entry. These fields are what make an alert actionable rather than decorative: a dispatcher reading the message in under a minute can verify the scene, read the dwell, and decide whether to talk down, dispatch, or log.

Cyrano taps the DVR's HDMI multiview output, the same composite signal that drives the guard monitor. That signal already has every camera on the recorder mosaiced into tiles, so one HDMI tap gives inference access to every feed at once. Install is under 2 minutes: HDMI in from the DVR, HDMI out to the monitor, network cable, power. It works on Hikvision, Dahua, Lorex, Swann, Uniview, and any DVR or NVR with an HDMI port. No ONVIF negotiation. No per-camera credentials. No camera firmware changes. Up to 25 camera tiles covered per unit, re-scoping to full-res on whichever camera the operator switches the DVR to fullscreen during an active incident.

Package theft at mailroom doors and lobby shelves, cable and copper theft at transformer pads and conduit chases, HVAC theft at condenser cages and line-set runs, cargo theft at loading-dock aprons and trailer yards, parking-lot theft including catalytic converter prep, and jobsite theft from conex boxes and staging areas. Each of these scenes shares the same structural pattern: a defined target zone, a pre-action pause while the actor positions, and a time window during which legitimate presence is zero or near zero. The filter ratio math works identically across all of them.

Because the channel staff already reads is the channel that stays read. WhatsApp is already on every property manager's phone for vendors, maintenance, and move-outs. Adding a single thread for theft alerts is a zero-friction change. Dedicated monitoring dashboards start around 250 dollars per camera per month because they assume a human watching a feed. One WhatsApp thread per property, with a tile thumbnail and a timestamp on every delivered alert, matches the communication pattern that Class B and C multifamily, commercial, and construction teams already use. It is a deliberate product choice tied to the filter-ratio thesis: the compression ratio only matters if the channel carrying the output is read.