Your smart home alerts are loud because the filter is reading 1 input. A real filter reads 6.

Smart home security alert filtering is the layer that decides which detected events become a notification on your phone and which get logged silently or dropped. The detector says 'a person was seen at 23:14 at the side gate, confidence 0.91'. The filter then asks a separate question: should this interrupt anyone? A real filter reads six inputs to answer it (class, zone, time, dwell, badge, posture). Most consumer smart home apps only let you tune one or two of those, which is why turning on every 'smart alert' setting in Ring, Nest, Arlo, Eufy, Wyze, or Blink still leaves you with a buzzing phone.

Because person detection is one input. A person at the front door at 17:30 (you, returning home) is the same input as a person at the side gate at 02:30 (worth checking). Until the filter also reads zone, time, dwell, badge, and posture, those two events score the same. Most consumer apps make the first the user's problem and ship the second one too. Ring's Smart Alerts (zones plus motion frequency), Nest's familiar faces (badge), and Eufy's AI human detection (class) each cover one corner of the six-input filter. None of them cover all six.

Nest covers the most for free users on Aware Plus: object class (person, vehicle, animal, package), familiar faces (badge), and zones. Ring's higher tiers add zone polygons and a motion frequency cooldown, which is a primitive form of dwell. Arlo's Secure plan adds package and vehicle classes plus zones. Eufy and Wyze ship class plus zones at the camera. Blink is the thinnest, with object class and a sensitivity slider. None of the six brands ship a posture input, and time-of-day rules are usually limited to schedules (camera on or off), not to per-zone class-time fire rules.

Posture is the geometry of the bounding box: walking through, leaning, crouching, hands on door, masked. It is computed by a small classifier on the bounding-box crop after the main detector fires. Posture matters because it is the difference between a person walking past your gate (drop) and a person leaning into the lock plate at the same gate at the same time of day (page). Consumer smart home apps do not currently expose posture. The closest is Ring's 'package theft' detection on the Pro stack, which is a narrow special case of posture (package being lifted from a porch), not a general posture input.

Dwell is how many seconds a tracked object stays inside a zone. Sensitivity is the threshold at which the motion detector fires at all. Lowering sensitivity drops detections; it does not let you say 'fire only if the person stays in the zone for more than 30 seconds'. Most break-in attempts have a long dwell at a target (gate, window, parcel shelf). Most legitimate visits have a short dwell. Filtering on dwell separates the two with one extra input. Sensitivity, set wrong, drops both. Ring's motion frequency is a cooldown that approximates dwell at the alert layer; it is not the same as filtering on the dwell of a tracked object.

Yes, on a self-hosted stack. A Frigate or Scrypted install with a Coral USB or Hailo edge accelerator and a small posture model can read all six. You configure zones in the Frigate config, time windows in Home Assistant automations, dwell from Frigate's tracked objects, badge from a smart lock or BLE presence, and posture from a custom inference plug-in. The trade-off is that you are now the integrator, the maintainer, and the on-call for the stack. For people with 4 to 6 cameras and time on weekends, this is the cleanest answer. For people with more than 8 cameras or with rentals to operate, the maintenance cost overtakes the savings.

Around 4 to 6 cameras for most apps, and around 8 cameras across the board. Single-doorbell setups (one Ring, one Nest Hello) are where consumer filters do their best work because the zone, the time, and the expected visitors are all narrow and stable. Adding a second camera fragments the rule set. By 8 cameras the operator is managing notifications across multiple zones with different time windows and different expected traffic, and the consumer apps stop trying. They will still send the notification; they will not help you decide whether the notification was worth seeing.

Two things change. First, the filter input set has to expand from 1-to-2 to all 6, because at this scale there is enough event volume that any missing input becomes operator-facing noise. Second, the architecture stops being 'app per camera brand' and starts being 'one event stream across the recorder you already have'. This is the gap Cyrano sits in. The unit plugs into the HDMI output of the DVR or NVR you already use, runs the 6-input classifier on every tile in the multiview, and emits one filtered event stream. It is not for a one-doorbell house. It is for a property with 8 to 25 working cameras and an operator who has stopped checking the app.

The class (person), the zone (rear gate), the time window (after_hours), the dwell duration (62 seconds), the badge state (none), and the posture (hands_on_door), plus the camera label, the property identifier, and a 10-second clip cropped to the triggering tile. That is enough for an operator to make a decision in under five seconds without opening the app. Compare that to a typical consumer app alert: 'Motion detected at Camera 2'. The same event, two different shapes. The difference is the five inputs that did not make it to the alert.

Probably not at the consumer tier, for two structural reasons. The first is that the audience for one-doorbell filtering does not feel the wall, so the willingness to pay for posture and badge inputs at that scale is thin. The second is that posture inference at the edge requires either an on-camera accelerator or a hub with one, and the BOM cost shifts the price to a tier consumers do not buy. Where the 6-input stack will actually ship is on commercial-grade boxes (the Cyrano category, and Verkada or Rhombus at higher price points) and on self-hosted stacks (Frigate, Scrypted, Home Assistant) that absorb the integration cost into the user's time.