Questions to Ask a CSI Sensing Vendor About Failure Before You Sign
Ask the vendor what their pipeline does when one access point drops out of the sensing session. Then notice whether the answer is about software or about physics. Channel State Information depends on several things at once: the transmitter's phase behaviour, the receiver's gain control, the firmware that exports the data and a baseline learned from an empty room. Each of these is a single point of failure unless the architecture was built to survive losing it, and a buyer with little lab time needs questions that bring each dependency into the open.
Who owns the CSI export
Many systems read subcarrier CSI through a chipset-specific firmware hook instead of the standardised measurement setup and report exchange defined in the WLAN sensing amendments, with 802.11bf as the reference point. So ask which interface they read from and what happened the last time the access point firmware was patched. The risk is the worst failure mode in this field because it makes no noise: an update changes amplitude scaling, subcarrier ordering or phase sanitisation, and the classifier keeps producing confident labels from corrupted input. Ask how they detect that. A credible answer covers schema validation on every report, a periodic reference capture against a known static scene and a drift alarm on the amplitude distribution. A vendor built on standardised report frames has a contract you can test against, while a vendor built on a patched driver is betting on a chipset roadmap neither of you controls.
Losing a link, or losing airtime
In a multistatic deployment, ask what happens to detection when one transmitter and receiver pair disappears. The risk is a model trained on a fixed link geometry. Remove one link and the input changes shape, so many end-to-end classifiers either reject the frame or pad it with zeros and quietly degrade. Designs that fuse per-link features late, with a confidence score for each link, let the remaining links outvote the gap and report reduced certainty rather than a wrong answer. Airtime fails in a subtler way. If the vendor specifies trigger-based sounding at a 10 ms interval, ask what the system does when data traffic on that access point stretches the interval or makes it irregular. Breathing and presence detection tolerate that well. Gesture recognition, which depends on micro-Doppler continuity, often does not, and the vendor should be able to tell you the interval at which they stop trusting a gesture output.
Ask where the empty-room baseline lives and who rebuilds it, because a system that recalibrates automatically will eventually learn a motionless person as furniture, while one that never recalibrates will drift into false alarms after the room is rearranged, so the vendor's answer tells you which of those two failures they decided to accept.
A demo that breaks things on purpose
Here is the decision rule for a shortlist with limited lab access: make fault injection part of the vendor demo or the acceptance test. Cut power to one access point mid-session, apply a firmware update to another and saturate a third link with bulk traffic. A system worth buying lowers its reported confidence within the same session, keeps working on the surviving links and recovers when you restore each fault, without retraining or a visit from the vendor's engineers. Where the vendor offers 5G-Advanced positioning as a fallback layer, make them show the handover too, because cellular and Wi-Fi measurements live in different timing domains and fusion logic that has never seen one source disappear tends to fail at exactly that moment. The trade-off remains even after a clean demo. Late fusion and conservative confidence gating usually give up some peak accuracy compared with a jointly trained model, so the real choice is between the best number in the vendor's demo room and predictable behaviour in month three.
Technical FAQ: Can I evaluate failure behaviour without running my own lab?
Partly. Ask for raw CSI logs and the matching classifier output from a recorded session in which the vendor removed a link and pushed a firmware update, then check whether the confidence values react at the moment of each fault. If the vendor cannot produce those logs, that tells you how much fault testing they have actually done.