paper-section-rewrite

Paper Section Rewrite v4

Takes one section of an academic paper and the LLM autonomously analyzes and improves it with a 4-Level critic loop, then outputs. No additional questions to the user. All analysis proceeds using only the input text, and the analysis results are included in the output for transparency.

Core design — 4-Level Critic Loop ★

Enforces the design order from top to bottom. Each Level proceeds to the next Level only after passing the CONVERGED gate— it does not touch sentences first.

Level 1 [내용]         의도·내용 1문장 + insight-first + 남는 그림 표  →  CONVERGED
Level 2 [문단 간 구성]  단락 outline + R1~R5 + 의도→방법→결과→해석 + lead 합=섹션 그림  →  CONVERGED
Level 3 [문단 내 구성]  단락별 카테고리 흐름 + 권장 흐름 일치  →  CONVERGED
Level 4 [문장]         단문 게이트 + 문장 연결 + 긍정형 우선 + 자연스러움 + 논리적 비약 + ML-beginner  →  CONVERGED

After all 4 Levels above are CONVERGED, the integration step:

[Phase 0]   입력 정리
[Phase 0.5] 장르 라벨 + 의무 단락 자동 도출
[Phase 0.7] 제약 원장 (어색함 → 규칙 → 재검사)
[Phase 1]   Level 1 critic loop — 내용
[Phase 2]   Level 2 critic loop — 문단 간 구성
[Phase 3]   Level 3 critic loop — 문단 내 구성 (단락별)
[Phase 4]   Level 4 critic loop — 문장 (단락별)
[Phase 5]   섹션 통합 critic + 자기 완결성 + 이력 독립성 + 그림·표 게이트 + 남는 그림 역검증 + 두 독자 평가(+ 논문 전체 유기성) + 저장본 재독
[Phase 6]   본문+부록 수식·notation 감사
[Phase 7]   LaTeX 빌드 검증

각 critic loop는 Up to 3 rounds or when the score stagnates terminate. Levels 1·2·3 enforce CONVERGED — it does not descend to the next Level with the structure not yet settled. If it cannot be settled within 3 rounds, report to the user only once and then retry, or mark only the weakest gate as cosmetic. Level 4 may proceed to the next Phase after marking remaining cosmetic issues (e.g., 1–2 naturalness patterns remaining).

Termination rule for the whole text. Separate from the 3 rounds per Level, the whole text counts the two reader evaluations of Phase 5 as one cycle. After passing all gates and the constraint ledger, one more cycle with 0 new criticisms must occur for it to end (up to 5 cycles). Even after the score stops, one or two new criticisms keep appearing, so the stopping signal is not score stagnation but 0 new criticisms.

When to use

• "Rewrite this section / restructure this section"
• "Settle the structure first and use the critic loop / in design order"
• "So that non-experts can understand / even for someone new to ML"
• "Conclusion-first / insight-centered / not a simple listing"
• "Without logical leaps / naturally / remove translationese / without AI vibes"
• "Make expressions like 'corresponds to / functions as' natural / reduce 'intend to' a bit"
• "Check mathematical notation in the body and appendix together"
• "Make the section self-contained"
• Copyediting a Korean/English academic paper section
• Copyediting one section of a Korean research proposal·proposal (necessity / prior research / method / utilization·expected effects / schedule / budget)

When NOT to use

• Adding margin comments to an entire paper PDF → paper-readability-review
• Writing step by step from the beginning while learning → academic-writing-trainer (written directly by the user, coached by Coach)
• Single-paragraph social-share summary → paper-digest
• General humanize → humanize-writing or humanize-korean

References

This skill references the following reference files during critic (shared with academic-writing-trainer).

• references/writing-principles-ko.md — 4 axes, 9 narration modes, recommended paragraph flow, short-sentence principle, naturalness principle, prose-style unpacking (ML + non-ML examples), fixing core terms and rotating conceptual terms, variety of sentence endings, self-contained principle
• references/genre-rubrics.md — mandatory paragraphs for 7 genres (research proposal full, including appendix) + recommended flow within the results section (intent → method → results → interpretation)
• references/feedback-corpus.md — review source text (R1–R5 few-shot across paragraphs + R6 sentence-level naturalness few-shot — research proposal worked example + R7 paper worked example + R8 loop case: turning awkwardness into rules, rereading the saved version)
• references/banned-phrases-ko.md / references/banned-phrases-en.md — prohibited words + grep patterns (including §7 copula substitutes, §8 cognate stacking, §9 deviation from the baseline register)
• references/scoring-rubrics.md — 4-axis scoring table + Pass/Fail gate + passing threshold

When outputting critic, quote the reference's R labels, category labels, and gate names verbatim.

---

Phase 0: Input organization

Automatically identify the following from the input text and context (do not ask the user):

• Section text (LaTeX or prose)
• Appendix text (if present — audited together with the main text in Phase 6)
• Page budget — if unspecified, maintaining the input length is the default
• Language — auto-detect KO/EN
• Available citation keys / math symbols / abbreviation dictionary (only those present in the input)
• Baseline register — Define in one line the style in which this text stands (e.g., "formal academic declarative", "research motivation writing mixed with first-person narration"). The register judgments in Step 4.i.2 and Step 5.1 look at whether it deviates from this baseline, not at a list (banned-phrases-ko.md §9)
• Constraints and fact notes given by the user — become the first lines of the Phase 0.7 constraint ledger and the Step 1.4c fact ledger

When explicit information is insufficient, apply reasonable defaults (no questions).

Phase 0.5: Automatic derivation of genre label + mandatory paragraphs

genre-rubrics.md See reference. Apply R1 and R2.

1. Automatically infer the section's genre from the input text (introduction / related work / method / results / discussion / appendix, etc.)
2. Retrieve the list of mandatory paragraphs for that genre
3. If it is a results/discussion section, internal recommended flow(intent (question) → method summary → results → interpretation) is also retrieved
4. Self-check whether the current section includes all mandatory paragraphs — if missing, create them in Phase 2 (R2 fail-fast)

---

Phase 0.7: Constraint ledger — the loop that turns awkwardness into rules ★

Gate lists catch only failures known in advance. In actual writing, awkwardness appears that no gate catches. This skill names the cause of that awkwardness, turns it into a rule, and reapplies the accumulated rules to the entire text on every cycle. This bundle of rules is called the constraint ledger. In actual loops, this loop changed the text more than polishing the style once did (feedback-corpus.md R8).

What goes into the ledger:

1. Constraints given by the user — such as length, prohibited words, number/link policy, fixed sentences, signature format
2. Sentences the user directly corrected — the before and after become examples of the rule as they are
3. Awkwardness the evaluator found but that did not fit an existing gate — comes from the list of awkward expressions in Step 5.6

The format of one line in the ledger (the two lines below are examples from actual writing and are not the default rule. In a paper, "inspired by" may be a normal expression):

Rule	Reason	전 → 후	Check
Do not speak of another's research as if it were the source of my research	The reader reads it as "they followed my work," and my contribution becomes smaller	"the step after your paper" → "As I understand it, your work asks …"	inspired by|modeled on|the step after|본떠|착안
Do not call a defined key term by another name	If you call it by another name, the reader reads it as a different concept	"trap … snare … hole" → fix as "trap," only reduce the count	grep for variants of the key term list

Diagnostic loop:

1. Extract awkward expressions sentence by sentence.
2. Check "why is it awkward" against the names of existing gates (copula substitutes, skipping, history dependence, noun-ending, etc.).
3. If a matching gate exists, fix it with that gate's prescription.
4. If no matching gate exists, create a new rule and enter it in the ledger. If you cannot write the reason in one sentence, do not turn it into a rule. A preference without a reason is not a rule but a taste.
5. From the next round on, reapply the entire ledger to the entire draft. New rules apply retroactively even to paragraphs that have already passed.

Operation:

• Keep the ledger as a file next to the document (e.g., constraints.md) so the next session can continue it.
• If a ledger rule and a general rule in references conflict, the ledger takes precedence. This is because it was decided for that document and that reader. Record the fact of the conflict in the report.
• Do not put rules in the ledger that delete facts or remove scope qualifiers (Step 4.i.7 safeguard).
• This step does not create questions. It only takes and accumulates the constraints and points the user has already given.

---

Phase 1: Level 1 critic loop — content

The LLM autonomously performs the same analysis as trainer Stages 1 and 2. Here, neither paragraphs nor sentences are touched. Only intent and content propositions are handled.

Step 1.1 — Automatic extraction of one sentence of intent

Summarize in one sentence what the input section is trying to do. Specify the verb (proposes / shows / analyzes / reports).

Step 1.2 — Automatic extraction of one content claim sentence

One sentence on what specific content achieves the intent.

Step 1.3 — Insight-first gate

paper-digest's core principle: No simple listing. "Why it matters" must come in the first sentence. Self-check whether the two sentences of intent and content are a simple listing of facts or insight-first.

의도 1문장 insight-first 점검:
  ✗ 단순 나열:  "본 연구는 모듈 A, B, C를 제안한다."
  ✓ insight:   "본 연구는 OOD 일반화 격차를 사전 지식 갱신으로 메우는 프레임워크를 제안한다."
                — 무엇이 새로운지(WHY)가 첫 문장에 들어옴.

Judgment: if the one intent/content sentence stays at "does what," ✗. If "why it is new / what it solves" is specified, ✓.

Step 1.4 — R5 alignment self-check

Self-check whether intent and content are consistent, and whether the thesis of the writing is the model or method itself rather than a tool.

When an R5 violation is found (e.g., if the intent is "proposing a model" but the content only emphasizes "Wasserstein distance"), make a note to rearrange it at the Level 2 outline stage.

Step 1.4b — The remaining picture table ★

What remains after reading. For each unit, write in one line the picture to leave with the reader. This criterion is not only for the abstract. It applies equally to the paper, sections, subsections, and paragraphs, and the picture of the upper unit must be the sum of the pictures of the lower units.

Unit	Where the remaining picture is placed	Relationship with lower units
Paper	Abstract	Connecting the section pictures makes the abstract
Section	Section introduction, 2–3 sentences	Connecting the one-line claims of the subsections makes the section picture
Subsection	One bold claim line (Finding)	The paragraphs each take on one of intent, hypothesis, method, result, or interpretation
Paragraph	First sentence (lead)	The remaining sentences only serve to support that lead
Sentence	One role	It starts from what the previous sentence finished saying

Procedure:

1. Before touching anything, make a table that writes in one line each the picture to leave for each section, subsection, and paragraph.
2. Check whether the one line of the upper unit matches the sum of the one lines of the lower units. If it does not match, it is not a sentence but a structure problem. At Level 2, fix the paragraph arrangement first.
3. Mark paragraphs and sentences that do not contribute to the picture as candidates for moving (to the appendix) or deletion.

Example. If a five-sentence abstract takes on purpose, approach, action result, internal result, and meaning one by one, then after reading, one picture remains with the reader: "autonomy increases risk and internal characteristics control it." If the same information is packed into eight sentences mixed with qualifiers and concessive clauses, it may be more accurate, but no picture remains (feedback-corpus.md R7-1).

Step 1.4c — Fact ledger ★

Decide what to claim before the sentences. Attach a status to each factual claim the writing makes.

Status	뜻	Handling
Confirmed	Confirmed through the input text, the user, or primary sources (paper abstract/body, personal page, data)	Write it
Needs confirmation	There is no source, or it relies on memory and estimation	Confirm before the final version, remove it, or reduce its scope
No result	Work in progress or with no result yet	Do not write it as a result. Write it only as an in-progress plan, or remove it

Rules:

• Descriptions of others' research (what they did and what they showed) are checked against the paper's abstract or body. The title, author order, and publication year are also subjects of comparison.
• Names, affiliations, and titles are checked against the source. The most common error is getting the surname right and inventing a given name.
• "Needs verification" is not quietly deleted. It is left as a list in the report so the user can judge.

Step 1.5 — Level 1 critic loop

Gate	Pass condition
Intent one-sentence clarity	Verb lock + 1 sentence
Content one-sentence consistency	Intent and supporting relationship
Insight-first	Rewriting from "listing what" to "stating why" completed
R5 thesis alignment	The method and results themselves, not the tool, are the thesis
Remaining figures/tables ★	There is one line per unit, and one line at the upper unit = the sum of the lines of the lower units
Fact ledger ★	Every factual claim has a status attached, and 0 sentences use "no results" as if it were a result

If it falls short, the LLM autonomously rewrites the two sentences (up to 3 rounds). Passing all 6 gates = Level 1 CONVERGED → proceed to Level 2.

---

Phase 2: Level 2 critic loop — inter-paragraph composition

How to arrange the paragraphsonly. Sentences are not touched yet.

Step 2.1 — Extract current paragraph leads

Extract the first sentence of each paragraph in the input section as its lead. N paragraphs, N leads.

Step 2.2 — Automatic labeling of paragraph roles

Automatically label each lead as one of the genre-mandatory paragraph roles.

Step 2.3 — R1~R5 self-check

R	Self-check	Action on violation
R1	Genre labeling completed (in Phase 0.5)	—
R2	Are all mandatory paragraphs filled in?	Add missing paragraphs
R3	Adjacent paragraph flow is normal (background→limitation→proposal→differentiation→evaluation)	Reorder paragraphs
R4	Semantic duplication between adjacent paragraphs	Merge the two paragraphs
R5	All leads support the section thesis	Remove or reassign surplus paragraphs

Step 2.4 — Check the recommended flow within the results section

When the genre is results/discussion/experimental report, an additional gate. genre-rubrics.mdthe internal flow of the results section:

의도(질문) → 방법 요약 → 결과 → 해석

Check whether these 4 stages map directly onto the paragraph lead sequence. Example:

P1 lead [의도/질문]: "본 실험은 ~를 검증한다."           ✓ 의도
P2 lead [방법 요약]: "ARC 1k 문제에 ~ 모델을 ~ 조건으로 평가한다."  ✓ 방법
P3 lead [결과]:     "본 모델은 zero-shot 78%로 baseline 대비 +12%p."  ✓ 결과
P4 lead [해석]:     "이 차이는 사전 지식 갱신 모듈의 역할로 설명된다."  ✓ 해석

Violation (e.g., results come first and intent comes later) → rearrange in Step 2.5.

Step 2.5 — Insight-first gate (paragraph lead)

각 lead가:

• ✗ Simple listing of facts ("This model consists of ~ modules") — only what
• ✓ Presenting insight ("This model closes the OOD gap through prior knowledge updating") — why is visible

Note violating leads for rewriting in Phase 3.

Step 2.6 — Level 2 critic loop (up to 3 rounds)

Gate	Pass conditions
R1·R2	Genre label + mandatory paragraph omissions 0
R3	Adjacent flow reversals 0
R4	Adjacent paragraph duplications 0
R5	All leads support the thesis
Internal flow within the Results section	Intent→Method→Results→Interpretation (if applicable to the genre)
Insight-first lead	All leads present an insight
Sum of leads = section picture ★	Reading only the leads in sequence yields the one-line section (subsection) from Step 1.4b
Connections between units ★	The final sentence of the preceding subsection/paragraph opens the question of the next subsection/paragraph. The same content is asserted in only one place, and other places take on the grounds or implications
Repetition of subsection shape ★	Subsections of the Results section repeat the same frame (one-line claim → design → results/interpretation → one table or figure)

All pass = Level 2 CONVERGED → proceed to Level 3.

---

Phase 3: Level 3 critic loop — within-paragraph composition (per paragraph)

Flow of sentence categories within a paragraphonly. Sentence expression is not touched yet.

Per-paragraph substep (for each paragraph P_i)

Step 3.i.1 — Labeling current sentence categories

Label every sentence in P_i as one of the 9 description modes (writing-principles-ko.md signal expressions used).

Step 3.i.2 — Compare with the recommended flow

writing-principles-ko.mdRefer to the recommended flow table by paragraph role. Identify missing mandatory categories / surplus avoidance categories.

Step 3.i.3 — Restructure the paragraph

Based on the above analysis, the LLM restructures P_i (add, remove, move, and reorder sentences). Polishing expression is done at the next Level.

Step 3.i.4 — Recheck labels after restructuring

Relabel to see whether the flow matches the recommended flow. Target role alignment 5/5.

Step 3 — Level 3 critic loop

Gate	Pass conditions
Category flow match	All paragraph role alignment ≥ 4/5
Mandatory categories satisfied	All mandatory categories appear per paragraph
Surplus avoidance removed	Avoidance categories 0 or appropriate for the paragraph role
Sum of sentence roles = paragraph picture ★	When the paragraph is covered, the lead comes to mind as is. Sentences that do not support the lead: 0
One picture per paragraph ★	If a paragraph has two or more results, split it. Method paragraphs keep only as much as needed to understand the results, and the rest goes to the appendix
Definition before use ★	Define a term at the place where it is first used, by explaining what it does

All pass = Level 3 CONVERGED → proceed to Level 4.

---

Phase 4: Level 4 critic loop — sentences (per paragraph)

Only here sentence expressionis touched. Four gates operate simultaneously.

Per-paragraph substep (for each paragraph P_i)

Step 4.i.1 — Single-Sentence Gate (Hard Rule) ★

Check whether every rewritten sentence has exactly one categoryonly. A sentence carrying two categories is immediately ✗ → split.

P_i 단문 게이트:
S1 [정의]              ✓ 단일 카테고리
S2 [혼합:비교+평가]    ✗ "X는 Y와 달리 Z를 더 잘 처리한다"
                          → 분할: "X는 Y와 다르다.[비교] X는 Z를 Y보다 더 잘 처리한다.[평가]"

Detection rule: -며 / -고 / -지만 / -에 의해connects two clauses, and if each clause belongs to a different category, split. Temporal·locative adverbial clauses·quoted clauses are recognized as single outside the category.

Length guide: KO 50 characters / EN within 20 words. Length alone is not ✗, but when length + two categories occur simultaneously, it is the highest priority.

Single-Sentence Ratio Gate (Hard): Within a paragraph, the ratio of single sentences (KO 50 characters / EN within 20w) is ≥ 60% enforced. If below, prose becomes heavy and readability for non-ML majors drops. Paragraphs below 60% automatically fail the substep → split or compress long sentences to pass the ratio, then relabel. 80% or above is praised.

Hybrid label [정의(+인과)] prohibited — unconditionally [혼합:A+B] ✗mark as and then split.

Details: writing-principles-ko.md the "Single-Sentence Principle — One Sentence, One Role" section.

Step 4.i.2 — Naturalness Gate

humanize-writing / humanize-koreandetection of core patterns. em-dash 0 is a quantitative gate, but naturalness is viewed as the accumulation of the following patterns.

Pattern	Korean signal	English signal
Translationese contrastive construction	"~에 대한 / ~에 의해 / ~을 통해"	"with respect to / by means of / through which"
Copula substitute idiom	"~에 해당한다 / ~에 위치한다 / ~로 기능한다 / ~로 자리매김하다 / ~방향에 있다 / ~핵심에 해당한다 / 입력으로 기능한다" (banned-phrases-ko.md §7)	overuse of "serves as / functions as / is positioned within / amounts to"
Cognate·redundant word stacking	"산출물을 산출 / 연구의 연구 / 다시 재구성 / 가능한 도달 가능한 / 동일한 X3" (banned-phrases-ko.md §8)	repeated cognate within a sentence
Mechanical parallelism	Unintended listing of "첫째 ~ 둘째 ~ 셋째"	Meaningless listing of "first ~ second ~ third"
Overuse of metacognitive expressions	"~라고 할 수 있다 / ~라고 볼 수 있다 / ~로 사료된다"	"it can be argued that / it is worth noting"
Excessive formality	"~함으로써 / ~에 있어 / ~에 다름 아니다"	"vis-à-vis / in light of / inasmuch as"
Noun-form accumulation	"~의 ~의 ~의" / "분석을 통한 모색"	nominalization chains
Register deviation	Suddenly deviating from the baseline register of Phase 0. If the standard is formal academic writing, written style ("아니한다 / 상기 / 전술한") ↔ colloquial style ("끌려간 / 흔들림 / 넓게 모은다 / 막고 / 가져온다") (banned-phrases-ko.md §9)	mixing archaic ("hereinabove") ↔ casual ("grab / a bunch of")
Em-dash overuse	— usage ≥ 1	— usage ≥ 1

Judgment: The above patterns per paragraph cumulative ≥ 3 = ✗. Rewrite as natural declarative sentences and connectives. (If two patterns appear simultaneously in one sentence, count only the more obvious one — no double counting.)

Rhythm (soft ⚠ only): If all sentences in a paragraph end with the same sentence-ending form (all "~한다") or are all similar in length, ⚠ note "monotonous endings/rhythm." The short-sentence ratio ≥ 60% Hard Rule can never be overridden. Handling: Split one or two long sentences into two short sentences (which also raises the short-sentence ratio) or naturally change the ending of a sentence. Not reflected in the score (writing-principles-ko.md see "sentence-ending diversity and sentence rhythm").

Step 4.i.3 — Logical leap gate

iterative-academic-writingCore principle of: the logical connection between adjacent sentences must be explicit. Leap detection patterns:

1. Conclusion leap — "It is X. Therefore it is Y." but the X→Y inference step is missing. (e.g., "loss dropped. Therefore generalization improved." ✗)
2. Implicit assumption — Presupposes a fact the reader does not know. E.g., "This is meaningful clustering because cosine similarity exceeds 0.8." → does not state why 0.8 is the threshold.
3. Numerical value → conclusion leap — "Accuracy is 78%. Our method is superior." → missing comparison target and statistical significance.
4. Tense confusion — "Existing methods cannot do X (present tense). Our method does X (past-tense result)." → mixed tenses blur the comparison.

Judgment: leaps per paragraph ≥ 1 = ✗. Add one sentence for the intermediate inference step or weaken with a hedging expression.

Step 4.i.4 — ML-beginner gate ★

Each paragraph's first sentence(lead) and paragraph body are checked for comprehensibility to an ML non-expert. 4 items.

1. Spell out abbreviations at first use: An abbreviation appearing for the first time in a paragraph without being spelled out → ✗
   • E.g., "RLHF is ..." (first appearance) → ✗ → "Reinforcement Learning from Human Feedback (RLHF) is ..."
2. Natural-language explanation before introducing formulas: A formula appears suddenly without natural-language explanation → ✗
   • E.g., suddenly \theta = \arg\min_\theta \mathbb{E}_x [\ell(f_\theta(x))] → ✗ → "This model learns parameters that minimize the average loss. That is, $\theta = \arg\min \ldots$" in that order
3. Four or more high-density jargon terms: Four or more ML jargon terms in one sentence (transformer, attention, KV-cache, FlashAttention, etc.) → ✗ → split into two or more sentences
4. Descriptive paraphrasing — no noun-form endings: Do not end defined terms or complex jargon in noun form; explain them with verbs. Instead of "X is Y (frozen and ended in noun form)," use "X does ~ to ~ (descriptive, paraphrasing how it works)."

Item 4 details — The most frequently caught anti-pattern:

Pattern	✗ Noun-form ending	✓ Descriptive paraphrasing
Solidified as a noun form after definition	"This method is the improvement of generalization ability through the effective use of metacognitive representations."	"This method uses metacognitive representations as a learning signal. This signal raises generalization even on out-of-distribution inputs."
Define jargon with jargon	"Self-distillation is self-knowledge distillation learning."	"Self-distillation is a way for a model to relearn its own outputs as if they were ground truth."
Noun phrase chain	"Improvement of OOD generalization performance through the introduction of a prior knowledge update mechanism"	"By introducing a prior knowledge update mechanism, it raises generalization performance even on out-of-distribution inputs."
Ending with an abstract noun	"This module is an implementation of trajectory matching."	"This module compares a person's solution trajectory and the model's solution trajectory step by step."

Detection signal (Korean):

• [가-힣]+의 [가-힣]+의 [가-힣]+ (three or more layers of noun phrases)
• ~을 통한 ~, ~에 의한 ~, ~의 ~화/~성/~력
• When a sentence ends with a "~이다 / ~다" noun-form ending and that noun is a jargon defined for the first time

Detection signal (English):

• noun phrase chain ≥ 3 (e.g., "the introduction of prior-update mechanism for OOD generalization improvement")
• nominalization (e.g., "the realization of X by means of Y" → "X realizes Y by ...")

Judgment: Check the above 4 items in both the lead and the body. A violation of ≥ 1 in even one item = ✗. Relabel after supplementing with paraphrases.

Step 4.i.5 — Quantitative grep + 4-axis critic

banned-phrases-ko.md Self-count with the regex of (or -en.md):

Korean thresholds:

• 것이다 / 것이며 / 수 있다 ≥ 3 per paragraph → expression axis -1
• 에 대한 / 에 의해 / 을 통해 ≥ 3 per paragraph → -1
• [가-힣]+적인 / [가-힣]+적으로 ≥ 3 → -1
• 해당한다 / 위치한다 / 기능한다 / 자리매김 / 방향에 있다 (copula substitute, §7) ≥ 3 per section → -1
• 하고자 한다 / 할 것이다 / 하려 한다 Ratio (§5.1): goja / 평서종결 > 0.5 → -1, target ≤ 0.3
• Nonstandard conceptual term (a 2+ character compound coined by the writer) ≥ 5 times in one section → rotation recommended, ≥ 8 times → -1 (rotation table = writing-principles-ko.md)
• Cognate stacking (산출물을 산출, 연구의 연구 etc., §8) ≥ 1 per sentence → ✗ split/replace
• 아니한다 / 상기 / 전술한 (written style) + 끌려간 / 흔들림 / 넓게 모은다 / 막고 (spoken style) combined ≥ 2 → register ⚠
• 됩니다 / 되어진다 / 지게 된다 0 recommended
• — (em-dash) 0 enforced

English thresholds:

• clearly / obviously / various / it is worth noting 0 recommended
• moreover / furthermore / additionally ≥ 2 → -1
• serves as / functions as / is positioned within / amounts to overuse → -1
• — 0

Step 4.i.6 — Sentence connection gate ★

Not sentence expression but between sentence and sentenceis examined. There are two failure types.

Failure	Shape	Prescription
Skipping	A reasoning link is missing. Example: "The directions are nearly orthogonal. The shared subspace is transferred." It does not say how they are shared if they are orthogonal	Write the missing link in one sentence ("The common component is small so it is not caught by cosine, but if the two tasks are combined they are separated")
Incoherence	Repeating the same thing (the same reservation three times), or lengthy method details unnecessary for understanding the result	Say it only once. Details go in the appendix

Connection rules:

1. The next sentence starts by taking the information (new information) that the previous sentence ended with as its subject or topic.
2. Write the chain of inference ("because", "so", "yet", "therefore", "on the other hand") into the text. Place connectives only where the logic turns.
3. Write parallel results in the same sentence pattern (Model A and Model B, addition and deletion).
4. The last sentence of a paragraph states the meaning of that result, and that meaning becomes the question of the next paragraph or subsection.
5. A pair that plays one role, such as cause and effect, is combined into one sentence. The short-sentence ratio has a lower bound of 60%, and **if it exceeds 80%, ⚠ "risk of note form"**. If four or more short sentences follow one another without connectives, ✗. (Example: "Editing is applied to the band. Gemma is L16–21. This is because a single layer did not move." → "Because the bet did not move when only one layer was modified, edit six consecutive layers (L16–21) together.")

Judgment: skips per paragraph ≥ 1 or incoherence ≥ 1 = ✗.

Step 4.i.7 — Positive-form-first gate ★

There are three kinds of sentences that appear negative. Do not handle them mixed together.

Type	뜻	Handling
Negative result	A result that contradicts the claim or fails to extend the claim under different conditions	In the main text, leave only a positive statement containing the scope("In Gemma, the target module steepens the gradient"). Put "It is not reproduced in LLaMA" in the appendix
Contrast that supports the claim	A null but part-of-the-claim result ("the readout direction does not move the stake" supports "reading and writing are different")	Keep it. Write it in positive form ("the readout direction reports the state, and the stake is determined by the action-derived direction")
Grammar-only negation	Content is positive but the sentence is negative ("~ cannot explain", "does not claim", "Not X:")	Change it to positive form

Threshold: one line of subsection claim and the negative form of the paragraph lead 0. Negative forms per paragraph ≤ 1 (excluding contrastive nulls). Detection patterns are banned-phrases-en.md §13, banned-phrases-ko.md §11.

Safeguard (Hard). When changing a negative result into a positive statement, (1) always retain the scope qualifier ("in Gemma", "in the three API models", "at L22"), and (2) check that it does not contradict the corresponding result in the appendix. If you erase the scope, this gate creates exaggeration. The same applies when deleting hedging expressions ("weak", "rather than accurate predictors"). Instead of deleting evaluative words, state the numbers and the controls that passed as they are. How far the claim goes is told by the numbers.

Step 4 — Level 4 critic loop (up to 3 rounds)

Gate	Pass condition
Short sentence (One Role)	Every sentence has a single category
Short-sentence ratio ★	Ratio of short sentences per paragraph (KO ≤ 50 characters / EN ≤ 20w) ≥ 60%
Naturalness	AI tell pattern accumulation per paragraph < 3 (including copula substitutes, cognate stacking, register deviation)
Copula substitutes	"constitutes / is located / functions / positions itself / lies in the direction" declarative sentences — 0 per paragraph (banned-phrases-ko.md §7)
Cognate stacking	"produce the output / research of research / reconfigure again"-type: 0 per sentence (§8)
Register consistency	0 sentences that suddenly deviate from the baseline register of Phase 0 (§9). In formal academic writing, deviations such as "does not" and "dragged along" are 0
Logical leaps	0 leaps per paragraph
ML-beginner	Lead and body pass all 4 items (expanding abbreviations / natural language before formulas / jargon ≤ 3 / Prose-style expansion — no nominalized endings)
Quantitative grep	Passes all thresholds (including the ratio of "it is that"·"about"·copula substitutes·"intend to ~" ≤ 0.3)
Sentence connection ★	0 skips per paragraph, 0 disjointedness. Short-sentence ratio 60–80%, 0 cases of 4 consecutive short sentences without connectives
Positive form first ★	0 negative forms in the one-line claim and lead, ≤ 1 negative form per paragraph. Preserve scope limiters + 0 contradictions with the appendix
4-axis integration	9/12 or higher

All passed = Level 4 CONVERGED. Proceed to the next paragraph for that paragraph. All paragraphs passed = Proceed to Phase 5. (Rhythm monotony ⚠ is cosmetic — it does not block proceeding to Phase 5.)

Paragraph substep output format

=== P_i ([역할]) Level 4 ===

[단문 게이트]      ✓ 모든 문장 단일 카테고리 / 단문 비율 67%
[자연스러움]       ✓ AI tell 1개 ("에 대한" 1회) — 임계(3) 미만
[계사 대용]        ⚠ S2 "맥락에 위치한다" 1건 → "이 흐름과 연결된다"
[동족어 stacking]  ✓ 0건
[register]        ✓ 문어체·구어체 이탈 0
[리듬]             ⚠ 종결 "~한다" 4연속 — 1문장 변형 권고 (cosmetic)
[논리적 비약]      ⚠ S3에서 수치→결론 비약 1건 (보완 필요)
[ML-beginner]     ⚠ S1 lead의 RLHF 약어 풀어쓰기 누락
[정량 grep]       것이다: 0 / 에 대한: 1 / 해당한다: 0 / 하고자 한다 비율: 0.20 / em-dash: 0  ✓
[4축]             통일 3 / 연결 3 / 완결 3 / 표현 2 = 11/12

조치:
- S2 "맥락에 위치한다" → "이 흐름과 연결된다" (계사 대용 §7)
- S1 lead에 "사람 피드백 강화 학습(RLHF)" first-use 풀어쓰기
- S3 수치 78% 다음에 baseline 대비·통계 유의성 한 문장 추가

다음 라운드 진행

---

Phase 5: Section integration critic + self-containedness gate

Step 5.1 — Section-level quantitative grep

Count banned-phrases regex across the entire section. What cannot be caught at the paragraph level (Phase 4) section-wide patternsare examined here:

• Accumulation of copula-substitute idioms (해당한다 / 위치한다 / 기능한다 / 자리매김 / 방향에 있다, §7) — section-wide ≥ 3 → ✗ → rewrite as declarative sentences
• "intend to ~" ratio (§5.1) — goja / 평서종결 > 0.5 → ✗ → convert more than half to declarative form
• Fixing key terms and rotating conceptual words — first distinguish key terms (concepts that the text defines and on which the argument rests) from ornamental noun phrases (bundles of modifiers the writer attaches). Do not rename key terms; if they are numerous, reduce only the count by using pronouns, omission, and sentence merging. If ornamental noun phrases occur ≥ 5 times, rotation is recommended; if ≥ 8 times → ✗ → writing-principles-ko.mddistribute using the rotation table in . (Standard academic terms may repeat — consistency matters more)
• Register consistency (§9) — total sentences that suddenly deviate from the baseline register set in Phase 0 ≥ 2 → ⚠ → return to the baseline register. In formal academic writing, literary style ("does not / the above / aforementioned") and colloquial style ("dragged along / wavering / gathered widely / block") are deviations
• Abbreviation after first occurrence — "auxiliary check not relying on an LLM" first appearance → thereafter "LLM-independent auxiliary check", "recall appearing on Korean public SNS" → thereafter "SNS recall", spelling out in full only at first appearance

Step 5.2 — Self-completeness gate

Check whether sections and paragraphs can be read without citing other sections. Grep for the following patterns:

Pattern	Korean	English
Section reference	Section [0-9], §[0-9], [0-9]장, 위에서 다룬, 앞서 언급한	Section [0-9]+, §[0-9]+, as discussed in, as we will see, aforementioned
Appendix reference	부록 [A-Z], appendix	Appendix [A-Z], App. [A-Z]
Forward reference	다음 섹션에서	in the next section, we will show

When detected, handle as follows:

1. If key definitions or equations are scattered throughout the body → insert a self-contained one-line summary paraphrased in this section
2. If dependence on the appendix is large → summarize the appendix's key results in one line in the body (no repetition, paraphrasing ✓)
3. Remove forward references where possible (keep only when necessary)

Verdict: section references 0개 recommended. 1–2 is ⚠. 3 or more means self-completeness ✗.

Step 5.2b — History-independence gate ★

Zero expressions that a reader of this text alone cannot understand. Whereas the self-completeness gate catches reliance on other sections, this gate catches reliance on previous versions and the work process. It applies to the body, captions, and appendix alike.

Type	English signal	Korean signal
Time markers	earlier, previously, now, no longer, still, originally, at the time	previous, earlier, now, no longer, yet, originally, at the time
Version or work names	corrected design, legacy runs, new/old version, main study, follow-up	corrected design, legacy, old version, main study, follow-up
Process descriptions	re-run, re-analysis, re-audit, we then/later, additionally ran	rerun, re-analysis, re-audit, additionally run
Editorial decision descriptions	the body keeps…, we moved…, is now reported in…, kept for transparency	the body retained ~, moved to ~, kept for transparency
Review history	in this revision, as requested, reviewers, rebuttal, camera-ready	in the revised manuscript, as requested, reviewer, rebuttal

There is one prescription. Do not call it by time; call it by content. If the distinction carries real information, keep the distinction and change only the name.

• "the corrected design" → "the matched-loss design" (a design in which the three options have the same expected loss)
• "legacy runs" → "the closed-model runs, whose mid-variance option loses slightly more"
• "three earlier single-layer protocols" → "three single-layer protocols at L22"
• "main study / follow-up" → the role of the experiment or the task name
• "re-analysis" → the name of the analysis ("a cluster-robust analysis")

Caution. The case where a work name hardens into a design name is the last to be discovered. If "corrected", attached while revising the manuscript, becomes the name of the experimental design in the body, the reader asks, "What was corrected and why, and were the results before the correction wrong?" A reference that merely points to the body ("Table 3 of the body") is not a violation. Statistical terms ("FDR-corrected", "Bonferroni-corrected") are also not violations.

Verdict: the above types even one instance remaining means ✗. Patterns are banned-phrases-en.md §12, banned-phrases-ko.md §10.

Step 5.3 — Cross-section duplicate-expression gate

Detect whether the same expression or sentence is repeated in different sections (when other sections are provided together in the input).

Action: when repetition is found, force paraphrasing. Each section takes a role—one defines, the other covers when to use it.

Exception: sentences entered in the constraint ledger as fixed sentencesare not changed. These are sentences deliberately kept identical when making multiple versions from one original (see "bundle mode" below).

Step 5.4 — Caption↔body duplication gate

Detect whether the body restates what figures and tables already contain. This is an axis that R4 (adjacent paragraphs) and Step 5.3 (across sections) cannot catch. It is the most common duplication in papers and the last to be found. This is because captions are written and revised separately from the body.

Procedure:

1. Extract all \caption of the section. Also read table cell values and figure axis labels, legends, and annotations.
2. For each figure or table, find and pair the body paragraphs that refer to it as \ref .
3. Judge each pair.

Judgment	Form
✓ Correct division of labor	The caption or table carries the details (values, conditions, legends), and the body gives only one line of insight
✗ Body repetition	The body rewrites the caption's explanation
✗ Duplicated numbers	The same numbers appear in both the table and the body
✗ Triplicated numbers	The same numbers appear in the table, body, and caption
✗ Row-by-row repetition	The body rewrites each row of the table as a separate paragraph

Judgment: if even one ✗ is found, the gate fails.

The remedy is not deletion but division of labor. What the table can say (what was measured and what the values are) stays in the table, and the body says only what the table cannot say.

• Why these items — basis for selection
• What the values mean — interpretation
• What could not be established — limitations

When removing numbers from the body, replace them with a reference pointing to the table. However, that number stays in the body where it is used as the basis for a design decision, such as contrasting two approaches or disclosing an exception. The goal is not to move values but to avoid saying the same thing twice.

Be especially wary of structures in which the rows of a table become the paragraphs of the body. If the table has N rows, the body also has N paragraphs, which also creates mechanical parallelism. In this case, group rows into meaning units to reduce the number of paragraphs.

Also prevent misreading in the opposite direction. A caption carrying the details and the body giving only a one-line summary is not duplication but a correct division of labor. If this is misjudged as duplication and the caption is emptied, the figure loses its self-containedness.

Step 5.4b — Figure·table gate ★

Figures and tables must be read as one unit with the body text. After the division of roles between captions and body text (5.4) is finished, look at placement and appearance.

Conditions for turning it on. Turn it on only when there are figures or tables and the output has pages (a manuscript built with LaTeX or PDF). If there are no figures or tables, or if it is a page-less document such as Docs or email, skip it and write "skipped" in the report. Style consistency and fixed color meanings can be checked first with the image files alone, and placement, margins, alignment, and readability are checked after building in Phase 7.

Item	Pass condition
Style consistency	Figures of the same kind have the same canvas size, font, font size, line thickness, and error bar shape. Choose one reference figure in the body text, and new figures follow it
Fixed color meaning	The same meaning is the same color in every figure (e.g., fixed conditions are green, variable conditions are red). Do not use colors without meaning
Nearby placement	A figure or table is on the same page as the paragraph that first mentions its content, or at the top of the immediately following page. LaTeX sending a figure to the top of the next page is allowed. If it is farther away than that, ✗
Left and right margins	For a table with a large empty width on the left and right, place two tables side by side or reduce the columns. Figures fill the body text width
Alignment	After inserting a new figure, the positions and sizes of other figures and tables have not shifted
Readability	The value is visible for each bar, the legend does not cover the data, and the text is readable at print size. Remove elements not needed for reading (such as reference bars)

Check with your eyes. Render the built PDF page by page and open the figures one by one. After fixing, render again and check whether other pages have shifted (Phase 7). Judgment: if there is even 1 case, ✗.

Step 5.5 — 4-axis integrated critic

scoring-rubrics.md Comprehensive scoring on 4 axes.

Item	Check
Readability	Understandable by an ML non-expert from the first sentence
Conclusion first	Is the first sentence of the paragraph the thesis
Insight delivery	Not a simple list, state "why it matters"
Hallucination	No facts or numbers not in the body text
Overclaim	Does the strength of the conclusion not exceed the evidence
Accumulation of nominalizations	No sentence with ≥ 3 nominalizations
Copula substitutes	"corresponds to / is located at / functions as / is positioned as / lies in the direction of" < 3 per section
Ratio of "intend to ~"	≤ 0.3 compared with declarative endings (>0.5 ✗)
Key terms and concept words	0 variants of key terms (adjust only the count), decorative noun phrases < 5 per section (8↑ ✗) — rotation applied
Register consistency	Total deviations from the reference register < 2
Remaining em-dashes	With grep — count == 0
Self-containedness	Section references ≤ 2
Duplicate expressions	No identical expressions with other sections (apply paraphrasing)
Caption↔body duplication ★	Body does not repeat numbers contained in tables/figures (division of roles)
Figure/table gate ★	Style consistency, fixed color meanings, close placement, left/right margins, alignment, readability (Step 5.4b, only for build manuscripts with figures/tables)
Transitions	Conjunction consistency between adjacent paragraphs
Consistency	Unified terminology and abbreviations
Balance	Paragraph length balance
Correspondence	Introduction↔conclusion correspondence

If NEEDS_WORK, apply the punch list and run another round. If CONVERGED, proceed to Phase 6.

---

Step 5.6 — Reverse verification of remaining figures + two-reader evaluation ★

Recheck the table from Step 1.4b from bottom to top.

1. Read the lead continuously. Extract only the first sentence of each paragraph and read them continuously. It must say the same thing as the one-line claim of the subsection. Reading only the one-line claims of the subsections continuously should yield the section figure, and reading the section figures continuously should yield the abstract. The abstract is a result that emerges at the end of this process, not a separate piece of writing.

2. First-time reader verification. Have a reader who has read only this text (a separate critic) read the unit, then write the remaining figure in one line. If the meaning differs from the pre-written one line, ✗. Do not give this reader previous versions, work records, or the author's intent notes.

3. Decision-maker reader evaluation. Read separately from the perspective of someone who makes a decision based on this text (an examiner, advisor, reviewer, or professor receiving the email). This evaluator produces four things.

   • Gist of reading only the lead continuously
   • One-line impression of the writer remaining within the first 30 seconds
   • Pass/fail for each gate and constraint ledger rule
   • List of awkward expressions and, for each item, one line on "why it is awkward"

   The last list goes into the Phase 0.7 diagnostic loop. The two evaluators do not see each other's results. The first-time reader looks at understanding, and the decision-maker reader looks at persuasion and awkwardness.

Judgment: If the reverse-verification one line of any unit conflicts with the one line from Step 1.4b, return that unit to Level 2. The decision-maker reader's points are returned to the relevant Level through the diagnostic loop. This one evaluation is one cycle of the entire text. When one cycle produces 0 new points, finish (maximum 5 cycles).

Step 5.6b — Whole-paper organicity ★

Turn this on only when the input is the entire paper (abstract, body, appendix). If only one section was received, skip it. Do not create a new loop. Use the two-reader evaluation and termination rules of Step 5.6 as they are, but have the readers read the entire paper rather than one section. There are four checks.

1. Abstract = sum of section figures. If you connect the section figures, you get the same statement as the sentence order of the abstract. Find sentences that are in the abstract but have no basis in the body, and things that are results of the body but missing from the abstract.
2. Connections between sections. The last sentence of the previous section opens the question of the next section. If, when the section changes, the reader asks "why is this story coming up," ✗.
3. One claim in one place. The same claim is made in only one place. The abstract handles the figure, the introduction the preview, the subsection the evidence, the summary the connection, the conclusion the implications, and the remaining slots handle only evidence or implications (writing-principles-ko.mdthe organicity of the three layers).
4. The interlocking of the body and the appendix. The pointer the body sends to the appendix actually reaches that content. The map at the head of the appendix matches the order of the body's subsections. If the evidence for a body claim is not in the appendix, or a claim not in the body is only in the appendix, ✗.

Numbers and notation are handled by Step 5.7, so they are not cross-checked again here. If the one line the reader wrote differs from the one-line paper (abstract) of Step 1.4b, return that section to Level 2.

Step 5.7 — Reread the saved version ★

All gates run against the draft in your head. But what reaches the reader is the saved result (file, document, email draft). If it goes through a tool or sub-agent, the draft and the saved version can differ. At the end, reread the saved version from the beginning, sentence by sentence. Skimming makes you miss things.

Item	What it catches	예
Non-words·typos	Words wrongly entered as similar characters, broken syllables	"dense map signal" → "damp map signal," "I would like to ask" → "I would like to ask," "asks" → "buries"
Proper nouns	Cross-check personal names·affiliations·paper titles·authors against the source	Cases where the surname is right and the given name is made up
Numbers	Cross-check body numbers and units against the source	Values in the table and body differ
Format constraints	The format rules of the constraint ledger	Em-dashes or links reappear during conversion
Mirror	If there are KO/EN versions, is the content the same paragraph by paragraph	Sentences fixed on only one side

Verdict: if there is even 1 case, ✗. After fixing, reread the saved version. Because this is a gate of facts and notation, not style, it must run at the end regardless of whether Level 4 was passed. Do not report completion before passing this gate.

---

Phase 6: Body+appendix equation·notation audit

Sweep across the body and appendix together. Audit across the entire text, not just a single section.

Step 6.1 — Symbol inventory

grep -oE "[a-zA-Z]_\{?[a-z0-9]+\}?"           # 첨자 있는 letter
grep -oE "\\\\(mathbf|mathcal|bm)\{[^}]+\}"   # 강조 letter
grep -oE "\\\\(ref|eqref|cite)\{[^}]+\}"      # cross-reference

Collect from the body and appendix respectively and create an inventory.

Step 6.2 — First-use definition verification

Is each symbol defined only where it first appears? If it appears in the body without being defined → first-use check in the appendix → if there is no appendix either, Orphan ✗.

Step 6.3 — Detection of duplicate symbol usage (mixed use)

Is the same letter not used with two meanings? E.g., in the body \theta = model parameter, in the appendix \theta = angle → ✗ → recommend changing the appendix symbol.

Step 6.4 — Subscript/formatting conventions

Item	Check
Subscript/superscript consistency	x_i vs x_{i} mixed use X
\mathbf vs \bm vs \boldsymbol	lock to one
Greek vs Latin mixed use	\beta vs bnotation of the same variable with X
Capitalization	\mathcal{D} vs \mathcal{d} consistency

Step 6.5 — Equation cross-reference integrity

\eqref{eq:foo} broken 0, \ref{sec:bar} broken 0, \cite{key} undefined 0.

Step 6.6 — Body vs Appendix consistency

Is the same definition not different in two places? If detected:

• Body defines, appendix extends → consistency OK
• Body and appendix give different definitions → ✗ → unify to the body definition + appendix cites the body

Step 6.7 — Detection of equation misuse (Misuse)

Equation mismatches the body meaning:

1. Error in the equation itself — dimension mismatch (e.g., matrix × vector dimension mismatch)
2. Mismatch between the body explanation and the equation — the body says "average" but the equation is a sum (missing normalization)
3. Unit mismatch — consistency of notation for token count, time (seconds/ms), probability (0–1 vs %)
4. Incorrect inequality direction — the body says "at least" but the equation is <

Step 6.8 — Detection of orphan equations

An equation appears without definition or reference. In the body, \eqref a numbered equation that appears without being referenced is ✗ (if it is never pointed to in the body, it is redundant).

Step 6.9 — Audit report output

=== Phase 6 감사 보고 ===

심볼 인벤토리: 본문 N개 / 부록 M개
  중복 사용 (혼용):  [\theta (body=param, app=angle) → 부록 심볼 변경 권고]
  Orphan:           [\xi 등장 1회, 정의 없음]
  Misuse:           [eq.(7) — 본문 "평균" 설명 vs 수식 sum, 정규화 누락]
  Cross-ref 깨짐:    [\eqref{eq:proof} 미정의]
  Body vs Appendix: [Loss 정의 부록과 다름 — 본문 정의로 통일 권고]
  단위 일관성:        [time을 ms와 s 혼용 — ms로 통일]

verdict: PASS / NEEDS_FIX

If NEEDS_FIX, revise the body and appendix, then re-audit.

---

Phase 7: LaTeX build verification

xelatex -interaction=nonstopmode main.tex && bibtex main && \
xelatex -interaction=nonstopmode main.tex && \
xelatex -interaction=nonstopmode main.tex

Verification items:

• Is the number of body pages within budget
• 0 remaining em-dashes
• 0 new warnings other than cosmetic warnings
• Render figures and tables page by page and visually confirm the alignment and readability from Step 5.4b (every time after adding or fixing a figure)
• Is the number of References pages appropriate

Output Format

The final output of this skill is revised sections + 4-Level transparency analysis report.

=== Section Rewrite Output (v4) ===

[Phase 0.7 — 제약 원장]
   규칙 수:           사용자 제약 4 / 사용자 수정 2 / 진단 고리에서 새로 생긴 규칙 1
   충돌:              원장 "끌려가다 허용" ↔ §9 격식 목록 → 원장 우선 (기준 register: 연구 동기 글)

[Level 1 Summary — 내용]
   섹션 의도 (1줄):    "..."
   섹션 내용 (1줄):    "..."
   Insight-first:     ✓ "what" → "why" 재작성 완료
   R5 thesis 정렬:     ✓
   사실 원장:          확인됨 12 / 확인 필요 1 (보고서 끝 목록) / 결과 없음 0
   Level 1 rounds:    1 → CONVERGED

[Level 2 Summary — 문단 간 구성]
   장르:               [방법 / 서론 / 결과 ...]
   의무 단락:          [동기 / 흐름 / 구성요소 / 예시]
   Outline Before:    P1 [???] / P2 [한계] / P3 [평가]
                      R 점검: R2 ✗ (제안 누락), R3 ✗ (역전)
   Outline After:     P1 [배경] / P2 [한계] / P3 [제안] / P4 [차별성] / P5 [평가]
                      R 점검: 모두 ✓
   결과 섹션 흐름:     의도→방법→결과→해석  ✓ (해당 시)
   Insight lead:      모든 lead가 insight 제시  ✓
   Level 2 rounds:    2 → CONVERGED

[Level 3 Summary — 문단 내 구성]
   P1 (배경): 정의 → 인과 → 예시 ✓ 의무 [정의, 인과] 충족
   P2 (한계): 비교/대조 → 분석 → 인과 ✓
   P3 (제안) [신설]: 정의 → 분석 → 과정 ✓
   P4 (차별성): 비교/대조 → 인과 ✓
   P5 (평가): 정의 → 과정 → 인과 ✓
   Level 3 rounds:    P3 2 rounds, 나머지 1 round → CONVERGED

[Level 4 Summary — 문장]
   단문 게이트:        ✓ 모든 문장 단일 카테고리 / 단문 비율 67%
   자연스러움:         ✓ AI tell 누적 단락당 평균 1.2 (임계 3 미만)
   계사 대용:          ✓ "해당한다/위치한다/기능한다" 0건 (단정문 재작성 2건)
   동족어 stacking:    ✓ 0건 ("산출물을 산출" → "구축" 1건)
   "~하고자 한다" 비율: ✓ 0.18 (≤ 0.3)
   register 일관성:    ✓ 문어체·구어체 이탈 0
   리듬:               ⚠ P2 종결 단조 ("~한다" 4연속) — 1문장 변형 권고
   논리적 비약:        ✓ 0건 (보완 후)
   ML-beginner:       ✓ 약어 first-use 풀어쓰기 적용
   정량 grep:         것이다 1 / 에 대한 2 / 명사형 2 / 해당한다 0 / em-dash 0
   Level 4 rounds:    1 → CONVERGED

[Phase 5 — 섹션 통합]
   자기 완결성:        ✓ 섹션 참조 1개 (허용 범위)
   섹션 간 중복:       0 (패러프레이징 1건 적용)
   캡션↔본문 중복:     0 (표 2 행별 반복 7문단 → 3문단 통합)
   그림·표 게이트:     ✓ 스타일·색 일치 / 그림 3이 두 쪽 떨어짐 → 해당 문단 쪽으로 이동 (빌드 후 재확인)
   개념어 과반복:      ✓ 최다 "분석 틀" 4회 (<5) — 회전 불필요
   register/계사대용:  ✓ 섹션 전역 통과
   4축 종합:          주제 8/9 / 내용 8/9 / 구성 11/12 / 표현 11/12 = 38/42
   두 독자 평가:       바퀴 3 (새 지적 4 → 1 → 0) → 종료
   논문 전체 유기성:   건너뜀 (섹션 하나 입력) / 논문 전체면: 초록=섹션 그림 합 ✓, 한 곳 한 주장 ✓, 본문↔부록 포인터 ✓
   저장본 재독:        오탈자 1 수정 ("묵는다고" → "묻는다고"), 고유명사·숫자 대조 ✓

[Phase 6 — 수식·notation 감사 (본문+부록)]
   심볼 혼용:         0
   Orphan:           0
   Misuse:           0 (보완: eq.(7) 정규화 추가)
   Cross-ref:        ✓
   Body vs App:       ✓
   단위:              ✓ (ms로 통일)

[Phase 7 — Build]
   Body 5.8 / 6.0 budget ✓
   신규 워닝 0

[Final Section]
   [수정된 LaTeX/prose]

[Changes summary (commit message)]
   "Restructure section: insight-first leads, 의도→방법→결과→해석 flow,
    add proposal paragraph, fix sentence-level leaps, RLHF first-use unfold,
    self-contain (drop §3.2 reference, paraphrase from appendix)"

Composition order (LLM autonomous enforced order)

1. Genre first — Phase 0.5, then constraint ledger Phase 0.7
2. Thesis next (Level 1) — Phase 1: intent + content + insight-first
3. Outline next (Level 2) — Phase 2: paragraph lead + R1–R5 + intent→method→result→interpretation
4. Paragraph flow (Level 3) — Phase 3: category flow per paragraph
5. Sentence last (Level 4) — Phase 4: short sentences + naturalness + leaps + ML-beginner
6. Integration — Phase 5: section integration + self-containedness + figure/table gate (if applicable) + two-reader evaluation (return to the diagnostic loop; include organicity if it is the whole paper) + reread the saved version
7. Math + Build — Phase 6 (main text + appendix) · Phase 7

There is no touching the sentences first. The LLM autonomously enforces the top→bottom order.

Iteration discipline

• Levels 1·2·3 enforce CONVERGED — skipping prohibited. Do not touch the sentences while the structure is not yet settled.
• Level 4 may proceed to Phase 5 after CONVERGED or remaining cosmetic markings
• Run the critic itself after every substantive rewrite
• Do not stop at WEAK ACCEPT (Levels 1·2·3 until CONVERGED)
• If it is a KO/EN pair of papers, after revising one side, mirror paragraph-by-paragraph
• The whole text ends with "one round that passes all gates and ledgers + zero new comments"(maximum 5 rounds). A score plateau is not a termination signal
• When a new rule enters the ledger, apply it again to paragraphs that have already passed
• Do not report completion before passing the saved-copy reread (Step 5.7)
• After inserting or fixing a figure or table, build and check Step 5.4b again. Fixing one figure can push the other side out of place
• If the input is an entire paper, run Step 5.6b together on every round of the two-reader evaluation
• No commit before passing LaTeX build verification
• Audit after every commit: 4-Level CONVERGED? insight-first? short-sentence ratio ≥ 60%? narrative-style paraphrase? self-contained? caption↔body duplication 0? formula mixing 0? R2 passed? matches the recommended paragraph flow? copula substitutes < 3? "~하고자 한다" ratio ≤ 0.3? concept-word over-repetition < 5? register consistent?

Bundle mode — when making multiple versions from one source

When making multiple versions from the same source, such as school-specific SOPs or audience-specific emails, do not run from the beginning for each version.

1. Finish the shared body text first in the order above.
2. In the constraint ledger, write down fixed sentences(sentences that go in identically in every version) and slots to change(sentences that differ per audience). Fixed sentences are not subject to the paraphrasing in Step 5.3.
3. Polish only the slots to change for each version at Levels 3·4. Verify the fact ledger separately for each version (the audience's paper, name, affiliation).
4. Compare the versions. If the closing sentences and questions of the slots to change overlap, ✗.
5. The saved-copy reread is done for each version.

The principle of not asking the user

This skill is an autonomous editing tool. Given only an input section, it performs all analysis internally and outputs a revised version + 4-Level analysis report.

Do not ask:

• "What is the genre?" — automatically inferred from the body text
• "Please tell me the intent in one line" — automatically extracted from the body text
• "From which paragraph should I polish?" — automatic processing per paragraph starting from Level 2
• "What length?" — default is to maintain the input length

Exceptions where you should ask (rare):

• The body text is too short or ambiguous to extract the intent
• Citation keys / formula symbol dictionary are needed but not present in the input
• No LaTeX build environment information

Only in this case is it an explicit question. Otherwise, proceed automatically with reasonable defaults.

Not asking does not mean ignoring the user's constraints. Constraints the user has already given, sentences the user directly corrected, and facts the user has confirmed are received and accumulated in the constraint ledger and fact ledger. Facts left as "needs confirmation" are passed to the list at the end of the report instead of being asked about.

Contrast: academic-writing-trainer는 Writing progressively while conversing with the user. This skill autonomously refines already-written text.