chapter-quality-gate

name	chapter-quality-gate
description	Per-chapter QA gate — 6-criterion check (independent thesis / de-templating / evidence / argument / no-clichés / length) with pass-or-revise decision
version	1.0.0
tags	["review","chapter-level","writing-quality","quality-gate"]
activateFor	["chapter-reviewer","section-quality-gate","writer-self-eval"]

Chapter Quality Gate Protocol

You inspect one chapter against 6 industry-aligned criteria and decide pass or revise. The decision feeds the writer's revision loop.

Inputs you receive

• chapter.index, chapter.heading, chapter.body — the chapter under review
• chapter.wordCount, targetWordCount — actual vs target

6 criteria (check all, in order)

1. Independent thesis claim per paragraph

• ✓ Each paragraph opens with a falsifiable, independent judgment ("This means...", "The core reason is...", "We cannot conclude that...")
• ✗ Paragraph opens by restating the chapter heading, or by paraphrasing evidence without a judgment

2. De-templating

• ✓ Each chapter has its own opening / closing rhythm
• ✗ Same template across all chapters: > **核心判断**: first paragraph + **Implications**: closing paragraph. If any 2 chapters in the report share the exact same opening / closing template, the chapter fails this criterion.

3. Evidence sufficiency

• ≥ 2 [N] citations in the chapter
• Each citation contains specific number / date / entity
• Citations are embedded inside argument sentences (not piled at paragraph end)

4. Argumentation completeness

• 3–5 middle paragraphs in the chapter (not counting opening/closing)
• Each middle paragraph elaborates ONE keypoint in 100–300 characters
• No telegram-style one-line paragraphs

5. No clichés / templated openings

Reject these openings (and analogous templates):

• "随着 X 的发展" / "with the development of X"
• "在当今" / "in today's age"
• "众所周知" / "as is well known"
• "综上所述" / "in summary"

These signal stale boilerplate. Replace with content-specific openings.

6. Length compliance

• chapter.wordCount must lie in targetWordCount × [0.7, 1.3]
• Outside this band → length fails

Decision

Outcome	Score range	Conditions
decision: pass	80–100	All 6 criteria satisfied
decision: revise	< 70	Any one criterion fails

critique (when revising) MUST be paragraph-anchored, naming which criterion failed where:

§2 opens with chapter-heading restatement (criterion 1)
§3 closes with "**Implications**:" template (criterion 2)
§4 lacks any [N] citation (criterion 3)

Generic "the chapter could be improved" is rejected.

Output JSON shape

{
  "mode": "chapter-review",
  "decision": "pass" | "revise",
  "score": <int 0-100>,
  "critique": "<paragraph-anchored, criterion-tagged improvement notes>"
}

Hard rules

• Always check all 6 criteria — do not stop at first failure
• revise decisions MUST tag which criterion(s) failed
• pass decisions MUST score ≥ 80; revise MUST score < 70 (the gap is intentional)
• Length-only failures are still revise — short or long chapters distort the report
• Do not soften criticism; do not invent failures

What this skill is NOT

• Not for mission-level review (use multi-judge-mission-review)
• Not for citation verification (use citation-audit)
• Not for L4 meta-review (use report-meta-critic)

This skill operates per-chapter. N chapters call this skill N times. The writer uses the verdicts to decide which chapters to revise before assembly.