top of page

Misattribution Has a Half-Life

Writer: DTEK Nerd
DTEK Nerd
Oct 15
2 min read

OBSERVATION 004

THE ASSUMPTION:  Misattribution is a binary state: attributed or unattributed.

At the moment an identifier changes, two records may look unrelated. Five minutes later, location continuity may suggest a connection. An hour later, behavior adds another clue. A day later, network or commercial observations add more. Given enough time, several weak signals can become a strong hypothesis.


That suggests a different way to think about misattribution.


Instead of treating it as a permanent state, treat it as a contest between separation and accumulating evidence. A protection can reduce confidence at T+0 without guaranteeing that confidence stays low.


The rate of recovery matters. So does the adversary. A database that relies on a stable identifier may lose the thread completely. An observer with continuous visibility, multiple data sources, or a long collection window may have more ways to rebuild it.


This is why point-in-time testing can be misleading. A system can look excellent immediately after an identity change because the test ends before correlation has time to work. Operationally, the more interesting question is what happens after hours, days, or repeated cycles of use.


We call this the half-life problem. Not because linkage confidence literally decays according to a neat equation, but because the metaphor forces the right question: how long does the separation remain useful?


That also changes how protections should be compared. Two techniques may both break an identifier successfully. One may remain difficult to reconnect over time. The other may produce enough recurring continuity that confidence returns quickly. Calling both simply “anonymous” hides the difference.


Misattribution is not only about creating separation. It is about sustaining uncertainty long enough for that separation to matter.


WHAT WE ARE WATCHING

Tests that measure whether linkage breaks at the moment of change, but never measure how quickly longitudinal evidence can put it back together.

bottom of page