Provenance, or it did not happen

Two kinds of explanation

Ask a Large Language Model why it believes something and it generates an explanation in the same pass as the claim. The explanation is plausible, often correct, and not a record. It is reconstructed, not retrieved.

A Large Universe Model builds beliefs by accumulating weighted, attributed observations, so the explanation is a log: these seven observations, in this order, with these weights, moved this belief from here to there on these dates.

Why this is structural rather than a feature

Provenance falls out of the architecture. Because a belief is built from observations rather than generated, the observations are already the representation. You do not add an audit trail; you decline to throw one away.

This is why provenance is native to the third generation and awkward to retrofit onto the first. There is nothing in a Large Language Model's forward pass corresponding to "the evidence for this."

What it buys

Deployability in regulated domains. Finance, medicine, law and safety-critical engineering all require reconstructable reasoning. A model whose conclusions cannot be unwound cannot be used, however accurate.

Correction propagation. When a source is retracted, you can find every belief that depended on it and re-derive. Without provenance, a retraction is untraceable and the error persists silently.

Source learning. Tracking which sources moved beliefs that later proved right is only possible if the attribution was kept. This is how evidence weighting improves rather than remaining a set of guessed constants.

Trust that survives being wrong. A system that is wrong and can show why it was wrong remains usable. One that is wrong opaquely does not.

The cost

Retention conflicts with compression, and provenance is usually the first thing sacrificed under storage or latency pressure. That sacrifice quietly converts a Large Universe Model into a stateful Large Language Model — confident, current, and unable to justify itself.