Decay and the honesty of forgetting

The stale-confidence problem

Systems that accumulate information tend to treat everything they have recorded as equally current. A note from March and an observation from this morning sit side by side with no marker distinguishing them, and the model reasons over both as though they were contemporaneous.

The result is a system that is confidently wrong about things that used to be true — arguably the worst failure mode available, because the confidence invites reliance.

Decay as a design decision

A Large Universe Model widens beliefs over time in the absence of confirming evidence. The rate is a property of the belief, not a global constant:

• A supplier's registered address decays very slowly
• A service's p99 latency decays over days
• A person's current priority decays over weeks
• A market's liquidity conditions decay over hours

Setting these rates is genuine domain modelling and is one of the places where a general system needs domain input.

Why this is not just forgetting

Decay does not delete. The observation remains, with its timestamp and provenance. What falls is the confidence that it still describes the present.

This distinction matters for audit: you can always reconstruct what was believed and why, including that the belief was based on evidence that had become old. A system that deleted stale data could not.

Decay makes absence informative

A useful side effect: when a belief has decayed to low confidence, that is itself reportable. Nothing has confirmed this in six weeks is often exactly what someone needs to hear, and it is invisible to any system that treats recorded facts as permanently true.

The tuning problem

Decay too fast and the model is perpetually uncertain and useless. Too slow and it is confidently stale. There is no general answer, and calibration against outcomes — did beliefs at this confidence level actually hold — is the only honest way to set it.