Training-Data Rights Ledger

Training-Data Rights Ledger A data-flow diagram generated by Archify. 01 / Sources 02 / Ledger (Layer 02) 03 / Admission 04 / Models 05 / Answers Licensed corpus · licence terms · 01 / Sources Licensed corpus licence terms Crawled content · reservation check · 01 / Sources Crawled content reservation check Internal and user data · legal basis · 01 / Sources Internal and user data legal basis Rights ledger · one row per source · 02 / Ledger (Layer 02) Rights ledger one row per source Corpus build gate · no row, no read · 03 / Admission Corpus build gate no row, no read Model versions · lineage to rows · 04 / Models Model versions lineage to rows Training summary · generated, not drafted · 05 / Answers Training summary generated, not drafted Withdrawal or order · affected models only · 05 / Answers Withdrawal or order affected models only record record record admit train disclose forward lineage Legend primary data data store data flow

The right to train, per source

  • • Acquisition channel, licence or legal basis, and the permitted uses
  • • For crawled content, the reservation check with its method and date
  • • Rows are written by the pipelines that acquire the data

Joined to lineage

  • • Every training run records the ledger rows it read
  • • Backward lineage says what trained a model; forward lineage says which models used a source
  • • Training-content summaries are queries over the ledger

When rights change

  • • A withdrawal, an erasure request or an order marks the affected rows
  • • Only the models that read those rows are retrained or retired