UWB is easy. Field calibration is not.

The system combines multiple radios: UWB for ranging and data exchange, GPS for outdoor positioning and time, and Bluetooth for local setup. Filtering and estimation mattered, but so did the operator path around them.

The radios have different authority and failure shapes. UWB contributes precise local timing and range evidence, GPS contributes outdoor position and a separate time reference, while Bluetooth provides a short-range provisioning channel. The architecture keeps those roles explicit so loss of GPS does not silently invalidate indoor ranging, and a convenient setup transport cannot become an implicit control-plane trust root.

A field product is never just the radio link. It is anchor geometry, calibration, clock behavior, deployment friction, battery constraints, enclosures, provisioning, and the tools that let someone see what the system believes.

Geometry controls observability. Anchors concentrated on one side of a tag can produce ranges that look individually plausible while the position covariance explodes along the weak axis. Non-line-of-sight paths add positive bias, metal and concrete reshape multipath, enclosure changes move antenna delay, and oscillator drift changes timestamp behavior. The estimator therefore needs residuals, covariance, outlier policy, and a degraded state—not only an x/y coordinate.

A calibration value is not an installation footnote. Antenna delay, anchor position, channel configuration, clock source, firmware build, and enclosure revision form a versioned input to every range and position estimate. If any of them changes without invalidating the old calibration, the system can become precisely wrong.

Calibration records belong to a device identity and deployment epoch, with explicit provenance and rollback. The operator should see which anchors are trusted, which are extrapolating, when residuals shifted, and whether a replacement device inherited incompatible parameters. This turns field calibration from a one-time ceremony into an observable lifecycle with reviewable state transitions.

Filtering must preserve uncertainty instead of laundering noisy ranges into a confident dot. The pipeline distinguishes rejected measurements, accepted ranges, predicted motion, solved position, and operator-visible quality. Timeouts and minimum geometry decide when the system should hold, coast, or declare the estimate unavailable.

A useful test matrix varies path blockage, anchor loss, sequence gaps, clock discontinuity, cold start, motion profile, and calibration mismatch. The goal is not one flattering median error. It is to prove bounded behavior when the radio evidence becomes contradictory and to make the reason for degradation available to the operator without revealing private site topology or field measurements.

The discipline transfers directly to real-time AI: noisy inputs are normal, partial failure is expected, and an invisible internal state is an operational defect.

Both systems turn probabilistic evidence into user-visible action. The implementation earns trust by naming uncertainty, preserving provenance, bounding stale state, correlating every stage, and providing an operator path when the model no longer supports the decision. The domain changes; the requirement to expose what the system believes and why does not.