Curriculum provenance and rights #

The curriculum has a broader source problem than the software corpus. A writing lesson may be excellent while being legally unavailable for model training. Source quality and source permission remain separate fields.

Source statuses #

allowed: project-authored or otherwise cleared for the listed uses.
allowed_with_attribution: content may be used only with the registered attribution and modification obligations.
reference_only: humans may consult it while designing independent material; source text must not enter prompts, targets, or generated datasets.
pending_review: the license permits some reuse, but treatment of datasets, model inputs, or weights has not been settled for this project.
excluded: the intended training use conflicts with the license or custody requirements.

Only the first two statuses may supply training or evaluation content.

Separate source lanes #

config/sources.toml remains the conservative software-documentation corpus: permissive software licenses, documentation and implementation in one pinned repository, and exact code evidence.

curriculum/sources.toml registers pedagogical material. Its records may teach writing choices without corresponding to a software repository. The two manifests must not share an implicit allowlist.

Original synthetic records #

Checked-in seeds use original synthetic passages and fictional product behavior. Their provenance names docwriter-original and states how they were created. A synthetic label does not excuse copying a published example with nouns changed.

CC BY 4.0 #

Google's covered developer-site content can be reused and adapted with attribution. A generated record derived from it must preserve:

exact page URL and retrieval date;
content hash or archived revision;
creator attribution required by the source;
CC BY 4.0 notice and link;
a statement that the material was modified when applicable;
separate license treatment for code samples and any third-party media.

Attribution belongs in dataset metadata and accompanying notices, not inside every model target unless the writing task itself requires it.

Diátaxis and the Red Hat guides are CC BY-SA 4.0. They remain pending_review because applying share-alike obligations to distributed datasets and trained checkpoints requires an explicit policy. Their ideas can be cited in human-authored curriculum documents; their source text cannot yet be ingested.

Open Technical Writing and Write the Docs use CC BY-NC-SA 4.0. They are excluded from the intended commercial-capable training path.

Copyrighted books #

The IBM handbook is a research reference. Its publicly listed nine quality characteristics help identify a useful field of concerns, but chapters, checklists, and before-and-after examples remain copyrighted. This curriculum states independently written operational criteria and does not reproduce the book.

Receipts #

Every bulk build must preserve source registry version, record-level source IDs, retrieved content hashes where relevant, derivation notes, generator identity, review results, and dataset digest. If a source cannot be reconstructed and its permissions cannot be shown, fail closed.