Every collision probability you have ever been shown rests on an assumption about uncertainty. In conjunction assessment, that assumption does more work than the orbit itself — and it is usually invisible.
The number that looks like knowledge
The public catalogue everyone screens against — the general perturbations elements published by the US Space Force and mirrored by CelesTrak — tells you where an object is believed to be. It says nothing about how well that is known. There is no covariance in the file. None.
This is awkward, because probability of collision is almost entirely a statement about uncertainty. The same two objects passing 800 metres apart can be a routine event or a serious one depending on whether the position error is tens of metres or several kilometres. Feed the same miss distance into two different uncertainty assumptions and you can move the probability by three orders of magnitude.
So when a tool shows you a clean figure like 1.4 × 10⁻⁴, a covariance was assumed somewhere. The question worth asking any vendor is a simple one: which one, and on what basis?
What we assume, and why we say it out loud
OrbitIQ models covariance from two things we actually know: how old the element set is, and what class of object it is. Uncertainty grows with the age of the observation, and it grows fastest along the direction of travel — in-track error dominates, which is why conjunction ellipsoids are long, thin cigars rather than spheres. Debris and spent rocket bodies get wider distributions than tracked, manoeuvring spacecraft, because they are harder to observe and less predictable.
That model is a reasonable engineering approximation. It is not the truth. So every assessment the platform returns carries the covariance it used and a caveat saying so, in the payload, not buried in a footnote.
Read the band, not the digits
The practical consequence is that you should treat the band — routine, elevated, high — as the signal, and the digits as an ordering device. A probability of 3.1 × 10⁻⁵ is not meaningfully different from 2.6 × 10⁻⁵. It is meaningfully different from 10⁻⁷.
And when a decision actually commits propellant, you should be holding the other operator’s conjunction data message, which carries a real covariance produced from their own tracking. Screening tools built on public data are for knowing where to look and what to ask about. We would rather be the tool that told you where the number came from than the one with the most confident-looking dashboard.
A Collision Probability Is Only as Honest as Its Covariance