This repository contains the source for an Internet-Draft defining a "provenance" extension for the Virtualized Conversation (vCon) standard. The extension gives generated content a named home for recording how it was produced by a generative model.
The primary document in this repository is:
- draft-howe-vcon-provenance: Defines a
provenanceparameter on the vCon analysis object (and on the dialog object, for machine-generated dialog) that records the model and provider, the decoding parameters (temperature, top_p, and the like), the prompt or a hash of it, the vCon elements supplied to the model as input, and a hash of the output. Because the record binds an output to its model, prompt, and inputs by hash, a signed vCon or a SCITT transparency receipt over it can attest to a verifiable derivation.
This document is intended for submission to the IETF vCon Working Group.
The extension is a Compatible vCon extension (Section 2.5 of
draft-ietf-vcon-vcon-core). It introduces no new top-level fields. It
registers a provenance parameter through the Analysis Object and
Dialog Object Parameter Names registries that the core specification
already provides, so generation metadata gets a real, named home rather
than being overloaded into the analysis body or schema.
It is the lightweight, single-call counterpart to the
Agent Session extension:
provenance describes one model call that produced one analysis or
dialog turn, while agent_session describes a whole autonomous run.
The two layer and share a common model-identity vocabulary.
Active work in progress. Feedback and contributions are welcome.
draft-howe-vcon-provenance.md: The source for the draft.lib/: Tooling for building and managing IETF drafts (auto-cloned).Makefile: Build commands.
make # build draft-howe-vcon-provenance.txt and .html
make next # build a versioned .txt for datatracker submission
Please see CONTRIBUTING.md for details on how to contribute.
- Initial creation of the repository and draft.