Observation, estimate, inference

I work on two things that look unrelated. One ranks satellite conjunction events by collision risk. The other tells commercial teams which properties in a territory are worth investigating.

They are the same problem. Both take remote observations of something nobody is going to go and physically inspect, produce a number, and hand it to someone who will act on it. And in both, the number is worthless unless the person receiving it can tell what kind of thing it is.

Four kinds of number that look identical in a table

Put these four side by side in a report and they render the same way:

• A direct observation — something visible in current imagery.
• A public record — a parcel record, a permit, a filing.
• A derived estimate — a usable-area calculation, a density measure, something computed from the above.
• An inference — a judgment connecting evidence to a likely conclusion.

They have completely different failure modes. An observation can be stale. A record can be accurate and describe a legal fiction. An estimate inherits every error in its inputs and adds its own. An inference can be perfectly reasonable and wrong.

Present them as a uniform column of values and you have destroyed the reader’s ability to distinguish those cases — while looking more rigorous, because uniform tables look rigorous.

The same distinction, in orbit

Conjunction data has the same tiers. The observed state of an object is one kind of thing. The risk estimate that ships with a conjunction message is a derived quantity computed under assumptions. A model’s predicted probability is a further inference on top of that. And the spread across an ensemble of such models is a statement about the model, not about the orbit.

BEACON keeps these apart deliberately. The predicted probability and the ensemble uncertainty are reported as separate quantities because they answer different questions. When the viewer renders an uncertainty envelope, it labels it a visual proxy, because the underlying number lives in probability space and not in kilometres.

Collapsing those layers would produce a cleaner interface and a dishonest one.

Confidence should route attention, not decorate

The useful version of a confidence score is not a badge. It is an instruction about where to spend the next hour.

A limitation that says “ownership unverified” is doing work: it tells a team that the next step is a records check, not a phone call. A limitation that says “data may be incomplete” is doing nothing at all. The first identifies a specific unknown and implies its remedy. The second is a disclaimer.

The same test applies to model uncertainty. Ensemble disagreement earns its place in BEACON because it routes events to human review and measurably catches most of the important ones. If it had only produced a number to display next to the prediction, it would not have been worth computing.

What remote sensing does not tell you

Overhead data is unreasonably good at some questions and structurally silent on others.

It can tell you what is physically there and how it has changed. It cannot tell you who owns it, who is authorized to make a decision about it, what contracts are attached to it, or what its condition is beneath the surface that faces the sensor. Those gaps are not resolution problems that a better sensor fixes. They are category errors — questions the measurement was never able to answer.

Most bad inference from imagery comes from crossing that line without noticing: reading an observable proxy as if it were the underlying fact. Roof age is visible; roof condition is partly inferred. Weather exposure is measurable; damage is not, and treating exposure as evidence of damage is exactly the sort of step that feels rigorous and is not.

The discipline transfers

Whether the subject is a conjunction or a warehouse:

• Label every value with what kind of evidence it is.
• Keep derived quantities visibly downstream of their inputs.
• Make confidence route attention to a specific next action.
• State which questions your data structurally cannot answer.
• Resist the uniform table. Rigour that comes from formatting is not rigour.

More on the research side is on the research page. The applied version, written for commercial teams, is on the Arden Meridian blog.