Validation and development status

Radar range figures routinely conflate three entirely different quantities into a single "range". At Ozense they are kept explicitly apart:

The gap between 30 m and 14–20 m is not a contradiction but the usual distance between a detection boundary without margin and a track range. A modelled range only becomes a track range once probability of detection, track promotion and track retention have been measured statistically against distance.

A link budget can compute an SNR-limited range from transmit power, antenna gain, bandwidth, integration time and noise figure. For the hardware in use, that calculation shows a physically plausible path to greater distances – under assumptions that have not yet been measured together.

Binding rule of language: range is a problem of aperture, dwell and integration, not a chip problem. But any figure in metres is worthless without stating target radar cross section, detection threshold and integration time. Datasheet ranges for automotive sensors refer to vehicle targets and must not be transferred to drones; the difference in backscatter is 25 to 35 dB, a range factor of roughly 4 to 6.

Specifically open:

• Whether the integration needed for longer observation times remains usable coherently, or fails on range and Doppler migration.
• Whether the assumed individual gains from coherent beamforming, longer dwell and track-before-detect actually behave additively – they have not been measured together.
• That longer observation times conflict directly with alert latency: a target at 15 m/s covers about 24 m in 1.6 s.
• That rain, a wet radome, radome insertion loss and enclosure and mounting effects on antenna pattern and phase are not yet quantified. Until then, all range forecasts are fair-weather laboratory assumptions.
• That flight elevation is a highly sensitive parameter: at 1.5 m radar height and 14 m range, an altitude of 5 m already costs about 5.8 dB two-way – and it was not recorded in any field session so far.

It follows that greater ranges are development forecasts, not product properties. No fixed product range is claimed.

A result is worth only as much as the procedure that produced it. Ozense therefore works with a pre-registered evaluation chain:

• Pre-registration. Thresholds and evaluation logic are fixed before the evaluation data is inspected. A gate is not readjusted after an independent holdout has been seen.
• Data separation. Development, validation and independent holdout data are kept strictly apart.
• Reproducibility. Raw replay and evaluation are version- and hash-bound, so a result can be regenerated from the same raw data.
• External ground truth. Target positions are checked against an independent reference, not against the system's own processing chain.
• Stage-wise evaluation. Detector, tracker, track validity and semantic assessment are reported separately rather than netted into one end metric.

This discipline is uncomfortable: it has repeatedly forced earlier internal results to be discarded as unsound. That is precisely its purpose.

• No confirmed drone identity. The output is a candidate, not a determination.
• No fixed product range. 14–20 m of field tracking is demonstrated; everything beyond that is calculated.
• No guaranteed field false-alarm rate. Uncertainty-calibration methods hold only under their own assumptions, and under domain shift those assumptions no longer apply.
• No blanket GDPR compliance. Local, event-driven processing eases the data protection assessment but does not replace it.
• No KRITIS certification, and no claim that operators are legally obliged to deploy a drone radar.
• No automatic frequency permission. Fixed-site operation in the 76–81 GHz band requires separate clearance; an application for research use is pending with the German Federal Network Agency.
• Not a counter-drone effector. Ozense detects, assesses, documents and alerts. No neutralisation.