Research Thread 01

Provenance as a product constraint

Building systems where every generated claim resolves to a retrievable source, and what that rules out at the architecture level.

The constraint is cheap to state and expensive to hold: every claim a system produces must resolve to something a person can open and read. Most of the interesting work is in what you cannot do once you have committed to it.

It rules out the fastest way to build one of these, which is to let a model assert whatever sounds right and leave the reader to sort it out. It also rules out summarising away the thing the summary was drawn from. Both products are built on it, and both are slower to build because of it.

The open question is where the line sits. A claim traced to a source that is itself unverified has moved the problem rather than solved it, which is why Recruiting.no separates confirmed from self-reported rather than treating a citation as the end of the argument.

02 Evaluation for judgement work 03 Small models inside a boundary