Human Behaviour

People’s preferences are not stable facts waiting to be retrieved. They shift with context, evolve over time, and often only become clear through exposure to concrete options. Systems that ask users to declare what they want at the start of a relationship begin with the least reliable data they will ever have.

Mechanism

The profile builds from what users do. Signals are weighted by the commitment they required:

1. a completed experience followed by a reaction carries the most weight;
2. A rejection followed by with a redirection, telling system where to go instead, carries significant weight;
3. A redirection without a prior rejection carries less;
4. browsing without action registers as weak negative signal.

Declared preferences set an initial direction but are treated as a prior, overridden progressively as behaviour accumulates.

In Orion , this produced a distinction between two types of preference that update at different speeds. :

Surface preferences (volume, price, distance, timing) update on a single confirmed signal. A user who sees three jazz results and redirects toward Something Quieter tells the system that the genre was alright, but the level of volume was wrong. The profile updates on volume immediately while keeping the genre intact.

Core preferences (genre, format, social mode) require the same signal to recur across sessions before the profile shifts. A user who rejects rock once may simply not be in the mood. When the user has rejected it across three sessions, the system updates that at the preference level.

The distinction between surface and core will manifest differently across domains, but the underlying principle remains.

What makes it distinct

Preference Accumulation is a relationship model, not a session model. Most systems optimise for the next interaction. This one optimises for accuracy across a relationship. The profile does not reset between sessions — each interaction either confirms or quietly revises what the system already holds.

Where it breaks

At the start, this model is weakest. Before enough signal has accumulated, the profile remains thin and the system falls back on other defaults to handle zero-states. It breaks when preferences shift faster than the signal decay can track: a user whose taste changes significantly will receive confidently wrong recommendations until enough behaviour builds up to align with the new information. In categories where people interact infrequently, that correction can take a very long time.