Temporal Coherence

An AI system can be individually correct in every decision it makes and collectively incoherent across them.

Picture a system making recommendations to a student over a semester. On Monday, it suggests a particular topic because the student seems ready for it. On Tuesday, seeing a weak response, it recommends stepping back to the prerequisite. On Wednesday, seeing a strong response, it recommends a more advanced topic. On Thursday, something in the scoring shifts and it recommends starting over from the basics. Each of these decisions, taken on its own, is defensible given the information available in that moment. Put together across a week, they trace out a path that no teacher would ever have walked, and a student travelling along it would reasonably conclude that the system has no idea what it is doing. The problem is not in any single decision. It is in the lack of a consistent story across them.

Temporal coherence is the discipline of making sure that decisions made over time hang together into something that makes sense as a trajectory. A good human teacher does this almost automatically, because they remember the shape of the last few lessons and are consciously trying to build on them. A system without memory, or with only shallow memory, is much more prone to swing wildly as the latest data point tugs it in a new direction. The fix is not to make each decision smarter. It is to design for coherence across decisions, so that the sequence of moves looks like a plan rather than a string of reactions. Most adaptive systems optimise for local correctness and pay no attention to global shape, and it shows. When you look back over a period where something or someone kept changing course in ways that made no sense, what was it that finally broke your trust in the process?

Last week we said the foundation matters more than the layer on top. This week we look at how a system can be right at every step and still tell a story that makes no sense across them.