Knowledge Provenance

What Provenance Must Capture

Provenance is not metadata about a file. It is the structured record of how a knowledge asset came to exist and how it has changed: which sources contributed, which computational activities acted upon it, which agents were responsible, and which downstream artifacts depend on it.

In AI-mediated environments that record must extend further, capturing model identity and version, computational parameters, retrieval sources, human instruction, automated agent identity, processing time, resulting output, and the dependencies that follow.

Preserving Relationships, Not Just Objects

The essential research insight is that value lies in the relationships rather than the artifacts. A final output treated as an independent object has already lost the information that makes it trustworthy.

Our work therefore models knowledge as a graph of entities, activities, and agents connected by typed relations, so that origin and transformation history can be reconstructed rather than asserted. This allows direct questions to be answered: where did this information originate, what datasets contributed, which models transformed it, and who interacted with it.

Decay and Loss

Provenance degrades. Custodians depart, systems are retired, context is lost, and undocumented transformations accumulate. We study that degradation as a measurable phenomenon rather than an inevitability, which allows institutions to detect where lineage is thinning before the knowledge becomes unusable.