| name | cap-article-writer |
| description | Use when writing or refreshing an evidence-backed CAP comparison article from a frozen writer-input.json, especially when the article must match the Dividend Golden CAP's decision value without allowing the model to invent facts, identities, numbers, or links. |
CAP Article Writer
Turn a frozen writer-input.json into a structured editorial.json. Treat the model as an editor, never as a fact source.
Workflow
- Read all of
writer-input.json before drafting.
- Read article-blueprint.md, evidence-language.md, and editorial-contract.md.
- Identify the reader decisions supported by the fact catalog. Keep Provider and Access Path scoped together.
- Draft
editorial.json only. Do not draft the final Markdown article.
- Cite one or more valid
fact_refs in every prose block.
- Select Provider × Access Path identities only through a supported
decision_type, structured recommended_access_path_ids, and access_path_id fields. The validator must be able to derive the same shortlist from frozen facts.
- Leave every number, date, digest, URL, price, latency, market fraction, Provider name, and chart path to the deterministic renderer.
- Validate the draft with the repository's article build or publication reproduce command.
- Ask the human/editorial gate to approve the prose digest in the Publication Profile. Do not treat this approval as empirical evidence.
- Compare the rendered article against golden-rubric.md. Revise and re-approve the editorial draft if any reader-value requirement is missing.
Non-negotiable boundaries
- Never infer support from Provider marketing or an unobserved market.
- Never merge Native and QVeris Access Paths.
- Never declare an overall winner or invent a composite score.
- Never turn
provider_negative into permanent non-support.
- Never turn
evidence_insufficient into failure or support.
- Never write quantitative literals, URLs, digests, or Provider names in editorial prose.
- Never read private raw evidence. Use only the frozen writer input.
The final article is assembled by code from locked facts, tables, charts, reproduction commands, and this constrained editorial layer.