Several independent kinds of evidence feed a topic reading. Some are about what people are searching for and talking about, some about where money and research are going, and some about what is already being built. Domainators reads its own curated list of sources, and which ones sit behind a topic, and what each one is worth, is not published.
They are not equally informative and they are not treated as though they were.
- Read every daySeveral independent kinds of evidence, each on its own schedule.
- Checked against its own historyA reading far outside it is quarantined and contributes nothing.
- Weighted by the recordEach kind counts for as much as its movement has preceded real outcomes.
Weights are learned, not chosen
Each signal type starts with a fallback weight, which is judgement, and exists only so that a fresh installation with an empty source registry does not score every topic at zero. Once forecasts start resolving, those weights are refit against the resolved record: a source whose movement preceded outcomes earns weight, and one whose movement did not loses it.
The refit runs against resolved forecasts only, and it never restates a forecast that has already been published. A learning engine that reaches back into its own archive is not keeping a record, it is editing one.
Timeliness is traded against certainty
Two kinds of evidence can both be right and still be worth different amounts, because they arrive at different points. Evidence that somebody has claimed a name is early and ambiguous. Evidence that somebody has built something and put it online is later and much harder to fake.
The engine is paid for lead time, so it does not simply prefer the more certain signal. It weighs each kind by how well its movement has actually preceded outcomes in the resolved record, which is a question with an answer rather than a matter of taste.
Quarantine
A source whose latest reading is far outside its own recent history is quarantined rather than fed into a forecast. The usual cause is not fraud, it is an upstream format change, an API returning a default, or a counter resetting, and all three produce a number that is perfectly well formed and completely wrong.
Quarantine is sticky on purpose. Degradation clears itself, because a transient outage should not need a person, but a quarantined source stays quarantined until somebody clears it by hand. A source that can re-admit itself after producing one absurd reading will re-admit itself after producing the next one.
A quarantined or dead source contributes exactly zero, and it does so on the board and in the forecast at the same moment, so the score you read and the opinion the forecaster acted on are never two different things.
Independence, and the difference between zero and unknown
Forty registrations from one operator is one opinion rather than forty, so a topic whose activity is concentrated in a single cluster is discounted. That requires knowing who registered what, which is not always possible.
So the discount is only applied when most of the observed activity could be attributed at all. Below that floor nothing is applied and the page says the figure is not available, rather than printing a zero. A reading nobody took is not a reading of zero, and treating absence as evidence is the failure this whole engine is built to avoid.
Where to look next
Reading the board puts these numbers back together into what you actually see on the trend board.