Installs fail at the survey, not the flash
By the time firmware misbehaves you can fix it remotely. What you cannot fix remotely: a sensor watching the wrong space, powered by a dying socket, on WiFi that drops nightly. One disciplined hour on site prevents all three.
The four questions, in order
1. What must the client be able to say? "Nobody entered cage A3 after hours" needs different placement than "the pump sounds healthy." Write the sentence first; it defines everything.
2. Where does the radio actually go? CSI measures the path between devices. Walk the space with a survey node streaming; watch amplitude while a colleague walks the target zone. If their movement barely dents the stream, move the node — do not trust the floor plan, trust the measurement.
3. Where is honest power? The #1 field failure is power, not software: cleaner-unplugged sockets, brownout-prone circuits, "temporary" extension cords. Demand a socket that nobody will touch, on a circuit the client names. Note the boot-reason telemetry from your sentinels — a fleet logging BROWNOUT resets is telling you which site has dirty power.
4. Which WiFi, exactly? 2.4 GHz only, get the SSID/PSK for that floor, ask about scheduled router reboots and captive portals. On enterprise WPA2 networks, request a PSK-based IoT SSID — do not burn hours fighting 802.1X from a microcontroller.
Deliverable: the placement sheet
For every device: name, MAC, exact location photo, socket ID, target zone, WiFi network. Register each one in your Latent device registry with the site name — this sheet is what turns "some boxes on walls" into an installation a colleague can service without calling you.