A camera can show many animals and still fail to answer the question a survey is supposed to answer. The failure is not that the picture is unclear. The failure is that “the number of saigas” was never one number.
The September 2026 manuscript, A Metrological Framework for Counting Saiga Antelope in Kazakhstan, treats that phrase as a measurand. SaigaVision is the field system built on the same chain.
Six quantities share one name
The population in a region, the animals available to the sensor, the animals inside the sample, the animals a model detects, the unique animals left after duplicate control, and the abundance estimate after a detection correction are not interchangeable. Two careful surveys can both be called a census and still be counting different things.
A visible total is a conditional observation. An abundance estimate is a statement about the population that could have produced that observation. Model confidence is not the inclusion probability of the survey design.
Unknown is a result
Adult males usually carry horns and adult females do not, but horn evidence disappears with distance, blur, occlusion, and head angle. The manuscript keeps a third label: unknown. That label is a measurement. Forcing every difficult animal into male or female writes a demographic bias into the record.
SaigaVision uses the same gate. Sex is assigned when the head is resolved. Otherwise the animal stays unknown, and the count does not pretend to be a sex ratio.
Height is not a convention
Megapixels do not state the ground sampling distance. Height, sensor width, focal length, and image width do. The manuscript’s reference geometry is explicit about the trade: a 0.20 m feature falls from about 33 pixels at 50 m to about 11 pixels at 150 m. Those figures check the equations. They are not a field accuracy.
Higher coverage does not preserve the nasal profile or the horns. A flight height copied from another species, or chosen because the map looks efficient, is not yet a measurement protocol.
What remains unmeasured
The manuscript separates three result levels. Analytical geometry can be replayed from the equations. The implementation record — calibration, grouped splits, hashes, and a replaceable detector — can be specified before a test is opened. Field detection, mask quality, sex coverage, track error, disturbance, and population abundance are not claimed, because the locked imagery and independent ground truth are not in the paper.
That boundary is the useful part. SaigaVision can record what was observed, where the coordinate uncertainty sits, and which claims are still blocked. It does not turn a bounding box into a census.