Why stacked signals lie

Take any company and look hard enough and you will find three recent changes. A page was rewritten. Someone was hired. A script appeared. Arrange them in a sentence and it reads like a transition.

This is what most intent scoring does, and it is why the output feels convincing and performs badly. A score built by counting matching signals cannot tell the difference between a sequence and a coincidence, because counting throws away the thing that carries the meaning: the order.

The fix is to score the ordering explicitly. We do it with four inputs, and the interesting one is order agreement: given the pattern we think is happening, did the events arrive in the order that pattern predicts? Events in the predicted order score high, the same events reversed score low.

Temporal clustering is second. A transition happens over weeks. Three changes spread across fourteen months are three unrelated changes, and three changes inside the same week are usually one person editing a website rather than a company changing direction. Both get penalised, at opposite ends.

Third is source diversity. If one source saw everything, you have one signal wearing three hats. Corroboration across independent sources is worth more than volume from one.

Fourth is type spread. Six job postings are not six signals. They are one hiring event, and padding the supporting set with them inflates confidence without adding evidence.

The practical consequence: a single dated fact, however strong, never produces a confident window here. One event caps coherence at 0.3, because one fact cannot imply a transition. That will feel conservative next to tools that will happily score a funding round at 90%. It is the difference between a rubric and an inference.