Capturing model identity, parameters, prompts, and retrieval sources so AI-mediated transformation remains attributable.
Our research is directed not toward evaluating the substantive correctness, truthfulness, or acceptability of information, but toward preserving the technical evidentiary structure necessary to determine origin, transformation history, attribution, and continuity.
This distinction matters. Systems that attempt to adjudicate meaning are contested and brittle. Systems that preserve the record of transformation are auditable and durable, and they leave judgment where it belongs, with the institution.
When a model blends training data, retrieved context, proprietary algorithms, and human instruction into a single output, conventional attribution fails. There is no discrete transfer to observe and often no record that any particular source contributed.
We study what must be captured at inference time: model identifier and version, computational parameters, retrieval sources, prompts and instructions, agent identity, authorization context, and downstream dependencies, for the resulting artifact to remain attributable later.
An adversary no longer needs to breach a secure network if it can subtly contaminate the data feeding it. As generated artifacts and autonomous systems become more capable, the integrity of the information flowing through a system becomes as consequential as the perimeter around it.
Authenticating knowledge, in that environment, becomes as fundamental as authenticating identity.
Which system, which version, under which parameters.
What sources were consulted and incorporated.
The prompts and authorizations that directed the work.
What later artifacts inherit from this output.