Disagreement should not be averaged

The failure

Two credible sources say incompatible things. The natural move is to weight them and produce a mean, which lands between two positions that nobody holds and that the evidence does not support.

If a reliable filing implies a supplier is solvent and an equally reliable litigation record implies otherwise, the answer is not moderately solvent. It is that the situation is bimodal and something is wrong with one of the sources.

Why this matters more in a continuous system

A Large Language Model asked to reconcile conflicting documents produces a fluent synthesis in one pass and forgets it. The damage is bounded by the conversation.

A Large Universe Model that averages conflicting evidence stores the average and builds on it. Every downstream belief inherits a position that no evidence supports, and the disagreement that would have flagged the problem is gone.

What to do instead

Keep the modes. A belief should be able to represent "either X or Y, not something between them."

Treat conflict as an event. Two streams that historically agreed and now do not is a first-class signal — often the earliest available sign that something has changed or that a source has degraded.

Track source divergence over time. A source that repeatedly conflicts with consensus and is later vindicated is valuable. One that repeatedly conflicts and is later wrong should lose weight. This is only learnable if disagreements are recorded rather than smoothed.

The design consequence

This argues against a single scalar confidence per belief and in favour of representations that can be multi-modal, plus explicit machinery for surfacing conflict rather than resolving it silently.

It is also the property most likely to be dropped under engineering pressure, because averaging is simple and honest disagreement is inconvenient to display.