| name | prose-verdict |
| description | Gates complete public-facing prose with a binary PUBLISH or HOLD verdict based on factual support, audience fit, clarity, structure, voice, and repetitive or generic writing patterns, without rewriting the text. Use immediately before publishing articles, essays, READMEs, landing-page copy, or announcements. Do not use for drafting, copyediting, private notes, or review of an incomplete excerpt. |
Prose Verdict
Read the complete publication target and emit a binary verdict. Never rewrite it.
Scope
Use immediately before publishing a complete article, essay, README, landing-page copy set, or announcement.
Do not use for drafting, "humanizing," polishing, copyediting, or judging an excerpt as if it were the whole work.
Required inputs
Require:
- exact file or text revision;
- intended audience and publication channel;
- title or visible promise;
- sources for factual, quantitative, temporal, and attributed claims;
- author-provided voice constraints;
- render when hierarchy, images, or captions matter.
Return HOLD when the target is incomplete or changed during review.
Read without rewriting
- Read once as the target reader before applying a checklist.
- State the promised value and whether the draft delivers it early enough.
- Locate every blocking finding precisely.
- Explain reader impact and required author action.
- Do not provide replacement sentences, a rewritten outline, or a polished alternative.
Run the publication gate
Use references/publication-rubric.md.
Check:
- factual and citation support;
- title/body agreement;
- causal scope and stated limits;
- contradictions and unresolved objections;
- reader load and delayed utility;
- section and paragraph density;
- voice against author constraints;
- repetitive contrast formulas, generic abstractions, empty intensifiers, stock metaphors, and mechanical summaries;
- privacy, attribution, copyright, and public-suitability risks;
- links, images, captions, and render integrity.
Do not use a readability score or AI detector as the verdict.
Run an optional scanner
If PROSE_VERDICT_SCANNER is configured, execute:
node scripts/run-scanner.mjs path/to/target
The adapter passes the target as a direct argument with shell evaluation disabled. Read references/scanner-adapter.md before using it.
A scanner pass is supporting evidence, not the publication verdict. A configured scanner failure produces HOLD. When no scanner is configured, continue with the rubric.
Decide
PUBLISH: no blocking findings, required evidence is present, the complete exact revision passes the rubric, and a configured scanner did not fail.HOLD: any blocking finding, missing evidence, incomplete target, stale revision, or configured scanner failure.
Nonblocking notes may accompany PUBLISH, but the verdict remains binary.
Stop and refuse
Refuse to:
- rewrite, polish, or make the prose "sound human";
- invent numbers, outcomes, quotations, feelings, people, or sources;
- certify a claim whose evidence is unavailable;
- judge a complete work from an excerpt;
- hide density or unsupported claims behind a clean scanner result;
- approve a new revision using an older verdict.
Output
Use references/output-contract.md. Include exact revision, blocker count, scanner state, blocking findings, nonblocking notes, evidence checked, and a freshness statement.
