Granular Provenance Metadata v0.1 replaces the binary “human-authored versus AI-generated” with a structured account of production practice.
Its first distinction separates AI-mediated work from AI-generated language. Mediation asks whether a model materially shaped the production chain. Generation asks whether model-produced wording remains in the final text. A work may be heavily mediated but linguistically rewritten, lightly directed but largely model-worded, or mixed at different stages.
Seven mediation types are defined: propositional, structural, linguistic, translational, research, editorial, and transformational. They may co-occur without being collapsed into an aggregate score.
Five attestation questions add process detail: who originated the propositions; what model language remains; who reviewed and ratified the final work; whether contribution seams remain recoverable; and who accepts legal, editorial, provenance, and reception responsibility.
The schema keeps the archive’s freeform substrate field and adds machine-readable declarations beside it. Values may be true, false, or undeclared. External platform categories can be recorded with reconciliation notes rather than silently imported as ground truth.
The design deliberately refuses categorical adjudication. It records what the depositor declares and leaves machine-facing distributional measurement to a separate layer.