@litfamily/litopencode 1.0.7 → 1.0.9

package/skills/lit-humanizer/references/ko-patterns-e-j.md

@@ -0,0 +1,387 @@
+# Korean pattern catalog: E–J
+
+Evidence labels follow the source’s own distinctions. Validated means that the source reports comparative human/model or task-matched evidence; weak means the cue is theoretical, genre-limited, or not directly measured for the exact item; hold means the source explicitly says not to use it as a rewrite basis. A validated result can support preservation. The metric-only indicators are review aids, not authorship scores.
+
+## E. Rhythm and sentence shape
+
+### E-1 | Low sentence-length variation
+- What: Sentences cluster at nearly the same length, often with no long sentence.
+- Why: Uniform length can make a page sound paced by a template.
+- Detect: Metric candidate; calculate sentence-length spread only after excluding headings, lists, and short slide labels.
+- Bad → better:
+> Bad: The valve is open. The pressure is stable. The pump is running. The line is ready.
+> Better: The valve is open and pressure is stable. The pump is running, so the line is ready.
+> Bad: The form is short. The form is clear. The form is easy to submit.
+> Better: The form takes two minutes to complete, and its labels are clear.
+- False positives: Procedures, slides, captions, and accessibility writing may deliberately use short sentences.
+- Evidence: weak; the source describes the rhythm cue but no comparable threshold for this item.
+
+### E-2 | Same sentence ending repeated
+- What: Several sentences end with the same Korean predicate form.
+- Why: Repeated endings can erase differences between observation, inference, and request.
+- Detect: Heuristic; inspect endings only within one register and genre.
+- Bad → better:
+> Bad: 담당자가 접수한다. 관리자가 검토한다. 팀장이 승인한다.
+> Better: 담당자가 접수하고 관리자가 검토한다. 팀장 승인은 마지막 단계다.
+> Bad: 이용자는 신청한다. 시스템은 결과를 저장한다. 운영자는 상태를 확인한다.
+> Better: 이용자가 신청하면 시스템이 결과를 저장한다. 운영자는 처리 상태를 확인한다.
+- False positives: Government notices, regulations, and checklist items often require consistent endings.
+- Evidence: weak; the source offers an editing metric, not a validated predictor for all Korean prose.
+
+### E-3 | Every paragraph has the same sentence count
+- What: Paragraphs repeatedly contain the same number of sentences.
+- Why: Identical blocks can signal automatic formatting rather than idea boundaries.
+- Detect: Metric candidate; compare at least four prose paragraphs, excluding list items and headings.
+- Bad → better:
+> Bad: Each paragraph has three sentences, regardless of the topic.
+> Better: Give the outage paragraph one sentence and the recovery paragraph the three details it needs.
+> Bad: Four sections each contain exactly two sentences.
+> Better: Keep two where sufficient and let the section with the test results carry its full explanation.
+- False positives: A legal template, article abstract, or two-sentence slide may require symmetry.
+- Evidence: weak; no source measurement is attached to this specific rule.
+
+### E-4 | All-short sentence rhythm
+- What: Nearly every sentence is short and structurally simple.
+- Why: A demand for concision can become a string of disconnected statements.
+- Detect: Judgment only; look for several simple clauses in sequence, not a low mean alone.
+- Bad → better:
+> Bad: 새 센서가 도착했다. 설치를 마쳤다. 수치가 안정됐다.
+> Better: 새 센서를 설치한 뒤 수치가 안정됐다.
+> Bad: 회의가 끝났다. 팀이 이동했다. 장비를 확인했다.
+> Better: 회의가 끝나자 팀은 장비를 확인하러 이동했다.
+- False positives: Preserve short sentences used for urgency, dialogue, and clear instructions.
+- Evidence: weak; the source adds this as a qualitative rhythm observation.
+
+### E-5 | Long comma-separated clauses
+- What: A sentence strings long clauses together with commas.
+- Why: Readers must hold too many dependent details before reaching the main verb.
+- Detect: Metric candidate; measure words between commas and inspect sentences with long clause spans.
+- Bad → better:
+> Bad: The pump arrived Monday, after customs released it, while the team was still waiting for the connector, which delayed the test.
+> Better: The pump arrived Monday after customs released it. The connector was still missing, so the test moved.
+> Bad: The office sent the notice, which listed the dates, before the portal update, while applicants continued using the old form.
+> Better: The office sent the notice with the dates before updating the portal. Applicants kept using the old form.
+- False positives: Long periodic sentences may be deliberate; retain them when their logic remains easy to follow.
+- Evidence: validated; the source reports a 1.97× difference in one essay corpus and warns this is the opposite of E-4.
+
+### E-6 | Many different comma boundaries
+- What: Commas appear after many different grammatical units in one text.
+- Why: Indiscriminate comma placement can make clause boundaries feel arbitrary.
+- Detect: Metric candidate; count comma-adjacent parts of speech only with a language-aware parser; otherwise mark judgment only.
+- Bad → better:
+> Bad: After review, the manager, on Tuesday, approved the form, and the vendor, later, sent it.
+> Better: The manager approved the form on Tuesday after review. The vendor sent it later.
+> Bad: 예산은, 4월에, 담당자가, 다시 검토했고, 부서는, 승인했다.
+> Better: 담당자가 4월에 예산을 다시 검토했고 부서가 승인했다.
+- False positives: The source restricts the signal to essays, news, blogs, Q&A, and reports; do not apply it to poetry or fiction.
+- Evidence: validated; the reported 2.44× separation is genre-bound.
+
+### E-7 | Register shift in addressee endings
+- What: A Korean document unexpectedly switches among levels of politeness.
+- Why: Endings encode the speaker-listener relationship, not just surface form.
+- Detect: Judgment only; compare paragraphs with the same audience and purpose.
+- Bad → better:
+> Bad: 안내해 드립니다. 다음 화면에서 눌러. 제출해 주십시오.
+> Better: 안내해 드립니다. 다음 화면에서 선택한 뒤 제출해 주세요.
+> Bad: 이 보고서는 결과를 설명한다. 검토해 주시기 바랍니다.
+> Better: 이 보고서는 결과를 설명합니다. 마지막 항목을 검토해 주시기 바랍니다.
+- False positives: Quoted dialogue, customer messages, and deliberate voice changes can switch registers.
+- Evidence: weak; the source labels this estimate and notes that the source study did not quantify it.
+
+## F. Repetition and modifiers
+
+### F-1 | Empty degree adverbs
+- What: Intensifiers are added without changing the measurable claim.
+- Why: They can make a report sound excited without giving a reason.
+- Detect: Judgment only; remove an adverb only if the sentence means the same thing without it.
+- Bad → better:
+> Bad: The patch is extremely useful.
+> Better: The patch removes one manual step.
+> Bad: The queue was very slow.
+> Better: The queue took 14 minutes to clear.
+- False positives: Keep emphasis that belongs to a speaker, quotation, or personal essay.
+- Evidence: weak; the source presents this as an editorial habit.
+
+### F-2 | Two modifiers with the same meaning
+- What: Adjacent adjectives or adverbs repeat one quality.
+- Why: The second word adds volume but no distinction.
+- Detect: Judgment only; check whether each modifier identifies a separate property.
+- Bad → better:
+> Bad: The room was quiet and silent.
+> Better: The room was silent.
+> Bad: The estimate is rough and approximate.
+> Better: The estimate is approximate.
+- False positives: Keep modifiers that set different dimensions, such as rough and preliminary.
+- Evidence: weak; no direct corpus result is reported.
+
+### F-3 | Function-and-role doublet
+- What: A noun phrase names both a function and its role when one covers the other.
+- Why: The doubled label lengthens the sentence without narrowing the meaning.
+- Detect: Judgment only; compare the two nouns and retain a real distinction.
+- Bad → better:
+> Bad: The board’s function and role is to set the schedule.
+> Better: The board sets the schedule.
+> Bad: Her responsibility and duty includes closing the lab.
+> Better: Her duty includes closing the lab.
+- False positives: Keep the pair when policy defines separate statutory functions and roles.
+- Evidence: weak; this is a concision cue, not a measured AI signal.
+
+### F-4 | Nominalization suffix chain
+- What: Abstract nouns accumulate derivational suffixes where verbs would be clearer.
+- Why: A sentence can become a stack of concepts with no visible actor.
+- Detect: Metric candidate; count suffix families at document level, but review each token for its part of speech.
+- Bad → better:
+> Bad: The implementation of the new policy improved service coordination.
+> Better: Implementing the new policy helped the teams coordinate service.
+> Bad: A structural modernization of the process began in May.
+> Better: The team began modernizing the process in May.
+- False positives: Keep established technical nouns, legal terms, and measured variables.
+- Evidence: weak; the source supplies a threshold but cautions that normalization cues need genre context.
+
+### F-5 | Repeated abstract “-적” compounds
+- What: Many abstract nouns are modified by a generic “-적” adjective.
+- Why: The pattern can make a paragraph sound formal while concealing concrete relations.
+- Detect: Heuristic; inspect three or more distinct compounds in one document.
+- Bad → better:
+> Bad: We reviewed the strategic, technical, and economic implications.
+> Better: We reviewed the strategy, system limits, and project cost.
+> Bad: The team needs a systematic approach to operational readiness.
+> Better: The team needs a checklist for opening the service.
+- False positives: Preserve conventional terms when the suffix creates a useful technical distinction.
+- Evidence: weak; the source gives a density rule without direct cross-genre validation.
+
+### F-7 | Generic policy verbs and abstract objects
+- What: Repeated general verbs such as expand, strengthen, or improve attach to vague abstract nouns.
+- Why: The reader cannot tell what action the plan requires.
+- Detect: Heuristic; pair the verb with its object and ask what someone will do.
+- Bad → better:
+> Bad: The city will strengthen its support for evening transit.
+> Better: The city will add two evening buses.
+> Bad: The team will improve its coordination structure.
+> Better: The two teams will share one release calendar.
+- False positives: Keep the verb when it is a formal budget or statutory category; “design” can be literal.
+- Evidence: validated; the source reports large corpus differences for several abstract nouns and actions.
+
+## G. Hedging and balance
+
+### G-1 | Repeated inference endings
+- What: Several statements end with the same “appears/seems to” form.
+- Why: Repetition can make different confidence levels sound identical.
+- Detect: Heuristic; vary syntax only if the source’s uncertainty remains exactly the same.
+- Bad → better:
+> Bad: 수요가 늘어난 것으로 보인다. 비용도 오른 것으로 보인다.
+> Better: 수요는 늘어난 듯하다. 비용은 작년보다 4% 올랐다.
+> Bad: The sample appears incomplete. The log appears delayed.
+> Better: The sample appears incomplete, and the log arrived a day late.
+- False positives: Do not change probability or negate an inference to improve rhythm.
+- Evidence: weak; source revisions focus on modality preservation rather than on a validated diagnostic threshold.
+
+### G-2 | Multiple qualifiers around one claim
+- What: Several words hedge one statement.
+- Why: Redundant qualifiers make it hard to tell what uncertainty remains.
+- Detect: Judgment only; preserve the one qualifier supported by the source.
+- Bad → better:
+> Bad: The result might possibly be somewhat higher.
+> Better: The result may be higher.
+> Bad: 비용이 다소 낮을 가능성이 있을 수도 있다.
+> Better: 비용이 낮을 수도 있다.
+- False positives: Medical, legal, policy, contract, and forecast language may encode distinct confidence levels; the source says to preserve it.
+- Evidence: weak; this is an editing suggestion with strict domain exclusions.
+
+### G-3 | Safe-balance vocabulary score
+- What: A document repeatedly uses both-sides and caution terms as a substitute for a decision.
+- Why: Repeated balance words can avoid naming a specific condition or tradeoff.
+- Detect: Lexicon metric exists, but do not use its count to justify edits.
+- Bad → better:
+> Bad: Both options have strengths, so a balanced approach is important.
+> Better: Option A costs less; Option B meets the Saturday service requirement.
+> Bad: 양쪽 의견을 모두 고려해 신중하게 접근해야 한다.
+> Better: 6월 전에는 주민 의견을 듣고, 예산안은 의회가 결정한다.
+- False positives: A genuine policy comparison may need balanced treatment.
+- Evidence: hold; source reports only five occurrences across 120 items and no document reached its proposed threshold.
+
+## H. Connectives and topic reference
+
+### H-1 | Sentence-initial connective density
+- What: Many sentences begin with the same transition words.
+- Why: Repeated connectors can give every paragraph an announced logic.
+- Detect: Metric candidate, but the source found a strong single-model effect; count by genre and never block.
+- Bad → better:
+> Bad: Also, the clinic closes at six. Therefore, patients must arrive early. Moreover, the desk stops taking names at five.
+> Better: The clinic closes at six, and the desk stops taking names at five. Patients should arrive earlier.
+> Bad: 또한 비용이 늘었다. 따라서 일정을 줄였다. 게다가 담당자도 바뀌었다.
+> Better: 비용이 늘어 일정을 줄였고 담당자도 바뀌었다.
+- False positives: A connector may mark a true contrast or consequence; the source found two models near the human baseline.
+- Evidence: weak; a high total came mostly from one model family.
+
+### H-2 | Alternating “but/however” synonyms
+- What: Consecutive paragraphs rotate between several equivalent contrast markers.
+- Why: The alternation can look like forced synonym variety.
+- Detect: Judgment only; check whether each transition has a distinct discourse function.
+- Bad → better:
+> Bad: The trial passed. However, the memory use was high. Nevertheless, it shipped.
+> Better: The trial passed, but memory use remained high; the team shipped it anyway.
+> Bad: 조건은 맞다. 하지만 비용이 크다. 그러나 일정은 짧다.
+> Better: 조건은 맞지만 비용이 크고 일정은 짧다.
+- False positives: Keep contrast markers where the argument genuinely changes direction.
+- Evidence: weak; the source gives an editorial pattern, not direct corpus validation.
+
+### H-3 | Repeated “this/that” topic pointer
+- What: Paragraphs repeatedly begin by pointing at the previous sentence.
+- Why: The topic pointer may hide the actual subject and make transitions vague.
+- Detect: Heuristic; check repeated forms inside one paragraph.
+- Bad → better:
+> Bad: This shows that the port is understaffed.
+> Better: The port lacks two night-shift workers.
+> Bad: 이 점에서 정책의 변화가 중요하다.
+> Better: 정책 변경으로 야간 신청도 가능해졌다.
+- False positives: The source says humans use this construction and only one model supplies much of the evidence.
+- Evidence: weak; model-specific count differences do not support a universal detector.
+
+### H-4 | Repeated redefinition marker
+- What: A paragraph repeatedly introduces a restatement with a “that is” marker.
+- Why: The marker can signal that the prior sentence did not say the same thing clearly.
+- Detect: Heuristic; remove only when the following words repeat rather than clarify.
+- Bad → better:
+> Bad: The cap is 20, that is, no more than 20 workers may join.
+> Better: No more than 20 workers may join.
+> Bad: 이 구간은 정체다. 즉, 차량이 멈춘 상태다.
+> Better: 이 구간에서는 차량이 멈춰 있다.
+- False positives: Keep a definition or explicit restatement when the audience needs it.
+- Evidence: weak; no direct comparison is reported for this entry.
+
+## I. Noun-heavy forms
+
+### I-1 | Repeated “it is that” ending
+- What: Several paragraphs close with an explanatory noun phrase ending.
+- Why: The form may place the conclusion after a generic wrapper.
+- Detect: Heuristic; inspect only repeated paragraph endings, not an individual sentence.
+- Bad → better:
+> Bad: The queue is long, and the reason is that two workers are absent.
+> Better: Two absent workers left the queue long.
+> Bad: 문제는 인력이 부족하다는 것이다.
+> Better: 인력이 부족하다.
+- False positives: The source found this form about twice as often in human Korean. Normal use should remain.
+- Evidence: validated; comparison rejects any blanket rule against the ending.
+
+### I-2 | Repeated dependent nouns
+- What: Several abstract sentence endings rely on nouns such as “point,” “fact,” or “case.”
+- Why: The writer may bury the actual predicate inside a noun phrase.
+- Detect: Heuristic; review repeated phrases such as “the point is that” rather than counting one noun.
+- Bad → better:
+> Bad: The key point is that the rate fell by 6%.
+> Better: The rate fell by 6%.
+> Bad: 확인해야 할 부분은 제출 시각이다.
+> Better: 제출 시각을 확인해야 한다.
+- False positives: Keep a dependent noun where it marks a real unit, choice, or case distinction.
+- Evidence: weak; no comparative evidence is attached to this broad word family.
+
+### I-3 | “The fact that” wrapper
+- What: A proposition is packaged as “the fact that” or “the meaning that.”
+- Why: The wrapper delays the assertion and can pad a conclusion.
+- Detect: Heuristic; rewrite only if direct statement preserves the relation.
+- Bad → better:
+> Bad: The fact that the key expired means the worker must sign in again.
+> Better: Because the key expired, the worker must sign in again.
+> Bad: 중요한 것은 일정이 이미 늦었다는 사실이다.
+> Better: 일정은 이미 늦었다.
+- False positives: Retain the noun when the proposition itself is being discussed or contrasted.
+- Evidence: weak; the source reports a model-skewed subvariant, not a general rule for all uses.
+
+### I-4 | Need-and-obligation report ending
+- What: A paragraph repeatedly closes with a policy recommendation or necessity claim.
+- Why: Identical endings can turn observations into stronger obligations.
+- Detect: Heuristic; count paragraph endings while comparing before-and-after obligation markers.
+- Bad → better:
+> Bad: The service should add weekend hours. The form should show the wait time.
+> Better: The service should add weekend hours before July. The form should show the wait time beside each slot.
+> Bad: 이용 절차를 단순화할 필요가 있다. 안내를 강화할 필요가 있다.
+> Better: 이용 절차를 단순화할 필요가 있다. 안내를 강화할 필요도 있다.
+- False positives: Never remove, merge, or move an obligation if that changes who must act or when.
+- Evidence: validated; the source reports a 7.22× task-matched difference and requires modality markers to remain intact.
+
+### I-5 | “Need” noun without an actor
+- What: A sentence names a need but not the person or action that can address it.
+- Why: The noun can obscure responsibility.
+- Detect: Heuristic; add an actor only when the source identifies one.
+- Bad → better:
+> Bad: A response is needed before Friday.
+> Better: The on-call lead must respond before Friday.
+> Bad: 추가 검토가 필요하다.
+> Better: 담당자가 표 4의 계산을 다시 검토해야 한다.
+- False positives: Do not invent an owner or strengthen a suggestion into an obligation.
+- Evidence: weak; the source’s treatment is an editorial caution against obligation inflation.
+
+### I-6 | Repeated abstract “ability” nouns
+- What: A document repeatedly describes what an entity can do using ability nouns.
+- Why: Abstract capability labels can replace the action itself.
+- Detect: Heuristic; inspect three or more repeated nominal forms.
+- Bad → better:
+> Bad: The tool has the ability to compare three files.
+> Better: The tool compares three files.
+> Bad: 팀의 대응 능력은 빠르게 향상됐다.
+> Better: 팀은 더 빨리 대응하게 됐다.
+- False positives: Keep capability names that are defined measures or formal evaluation dimensions.
+- Evidence: weak; no direct comparative count is reported for this item.
+
+### I-7 | Unnamed analyst conclusion
+- What: A sentence attributes a conclusion to an anonymous analysis or evaluation.
+- Why: The phrasing borrows authority while omitting who reached the judgment.
+- Detect: Heuristic; inspect the previous sentence for an explicit, nearby source.
+- Bad → better:
+> Bad: Demand is rising, according to the analysis.
+> Better: The March report says demand rose 8%.
+> Bad: 비용이 낮아졌다는 평가다.
+> Better: 연구팀은 비용이 낮아졌다고 평가했다.
+- False positives: Korean journalism may use the form after a named source in the preceding sentence.
+- Evidence: validated; the source found human 0 and model 5, all in interview or citation tasks, so genre and attribution context are essential.
+
+## J. Visual punctuation
+
+### J-1 | Bold on most body phrases
+- What: Bold styling marks many ordinary nouns or clauses.
+- Why: Repeated emphasis removes the visual hierarchy.
+- Detect: Heuristic; inspect body prose separately from navigation labels.
+- Bad → better:
+> Bad: The team **reviewed every item** and **approved each change**.
+> Better: The team reviewed every item and approved each change.
+> Bad: **Deadline:** Friday. **Owner:** Mina. **Status:** Open.
+> Better: The review closes Friday. Mina owns it, and the item remains open.
+- False positives: Keep bold labels in forms, accessible scanning aids, and slide hierarchy.
+- Evidence: weak; no Korean-specific comparison is given.
+
+### J-2 | Quotation marks used for emphasis
+- What: Ordinary concepts are placed in quotes despite not being quoted speech or disputed language.
+- Why: Repeated scare quotes make the writer appear to distance themselves from their terms.
+- Detect: Heuristic; count emphasis quotes, not punctuation inside a quotation.
+- Bad → better:
+> Bad: The “quick” review took two days.
+> Better: The quick review took two days.
+> Bad: The team called the result “progress” after the outage.
+> Better: The team called the result progress after the outage.
+- False positives: Preserve direct quotations, exact labels, and a term under linguistic analysis.
+- Evidence: validated; source reports 2.4× more emphasis quotes in task-matched AI text, while noting task dependence.
+
+### J-3 | Decorative em-dash clusters
+- What: Em dashes repeatedly insert dramatic side comments or clause pivots.
+- Why: The punctuation can give every sentence the same staged pause.
+- Detect: Heuristic; inspect repeated decorative use, not a single dash.
+- Bad → better:
+> Bad: The review—one day—then build—two more days—finished Friday.
+> Better: The review took one day. The build took two more, and both finished Friday.
+> Bad: The team—after a short delay—sent the file—before lunch.
+> Better: After a short delay, the team sent the file before lunch.
+- False positives: Preserve dashes already present in the source, dialogue, and an author’s established voice.
+- Evidence: weak; the source’s current guidance limits the cue to repeated decoration.
+
+### J-4 | Parenthetical afterthought chain
+- What: Many parenthetical remarks interrupt otherwise complete sentences.
+- Why: The aside can make the main claim hard to follow.
+- Detect: Judgment only; check whether the aside belongs in a sentence, footnote, or note.
+- Bad → better:
+> Bad: The trial (which began in June) (after a short delay) ended in August.
+> Better: The trial began in June after a short delay and ended in August.
+> Bad: The owner (Mina) (the night-shift lead) signed.
+> Better: Night-shift lead Mina signed.
+- False positives: Keep technical expansions, citations, and a parenthetical that carries the author’s voice.
+- Evidence: weak; the source describes overuse but does not mark this ID as a confirmed signal.

package/skills/lit-humanizer/references/ko-patterns.md

@@ -0,0 +1,31 @@
+# Korean patterns
+
+This is the routing page for the full 85-entry Korean catalog adapted from the pinned im-not-ai taxonomy and the local D4 research synthesis.
+
+- A–D: references/ko-patterns-a-d.md
+- E–J: references/ko-patterns-e-j.md
+- Measurable, source-tested signals and threshold caveats: references/ko-metrics.md
+- Script and golden fixtures: scripts/ko-metrics.mjs and fixtures/ko-metrics-golden.json
+- Golden inputs and expected metric ranges: fixtures/ko-metrics-golden.json
+- Detector cues and current tiers: ../rules.json
+- Reader-facing placement and all 26 real-line rewrites: taxonomy.md and deliverable-channels.md
+
+## Evidence labels
+
+Each entry records the source taxonomy’s evidence status in one of three forms:
+
+- validated: the source describes a comparative corpus result, task-matched result, or measured review set. The entry preserves the source’s genre, model, sample-size, and direction caveats.
+- weak: the source offers a linguistic or editorial hypothesis without a direct comparison for that exact cue, or warns that the result depends on model or genre.
+- hold: the source explicitly says not to use the cue as a rewrite basis yet.
+
+Validated does not mean “remove this phrase.” The source has several findings where people use a form more often than models. Those entries say so and recommend preservation. A validated measurement is evidence about the measured sample, not a universal authorship rule.
+
+## ID inventory
+
+A: A-1–A-24. B: B-1–B-4. C: C-1–C-12. D: D-1–D-14. E: E-1–E-7. F: F-1–F-5 and F-7; the source has no F-6. G: G-1–G-3. H: H-1–H-4. I: I-1–I-7. J: J-1–J-4. Total: 85 IDs.
+
+Each entry contains an explanation, reason, detection method, two bad-to-better pairs, false-positive guidance, and the source’s evidence-strength label. Pattern entries are compact by design; consult the original read-only taxonomy for its bibliography and technical history. The port uses fresh examples and wording.
+
+## Safe use
+
+Read A-1, A-2, A-10, A-11, A-16, A-17, A-20, A-21, A-23, A-24, C-1, C-8, E-5, E-6, E-7, G-1 through G-3, H-1, H-3, I-1, I-4, I-5, I-7, and J-2/J-3 especially carefully. These entries contain preservation gates, dataset caveats, held status, modality restrictions, or genre limits. Do not count a phrase and rewrite automatically.

package/skills/lit-humanizer/references/rewrite-playbook.md

@@ -0,0 +1,54 @@
+# Rewrite playbook
+
+Use this procedure when a user asks for a rewrite or a deliverable needs a final editorial pass. It aims for a natural, accurate result, not a low detector score at any cost.
+
+## 1. Mark
+
+Read the whole section before changing a sentence. Identify its intended reader, format, voice, and purpose. Mark likely drafting residue separately from valid technical terms, citations, quotations, and limits. A warning is a prompt to inspect the passage; do not mechanically rewrite it.
+
+For substantive work, keep a temporary fact table outside the final deliverable:
+
+- claim or instruction;
+- exact number, unit, date, and proper name;
+- source or quotation;
+- modality and uncertainty;
+- condition or scope boundary;
+- where it appears in the revised draft.
+
+The table is a working aid. Do not emit it unless the user asked for it or it belongs in an internal record.
+
+## 2. Rewrite
+
+State the useful point directly. Prefer the actor and verb when they are known. Keep a sentence if its structure is natural for the field or the writer. Use bullets for parallel items or actions; use prose for a line of reasoning. Vary rhythm in response to the content rather than to a formula.
+
+Remove a phrase only if the next version preserves its function. If a sentence contributes no fact, reasoning, or voice, cut it. If it contributes meaning, rewrite it with the same meaning. Do not add anecdotes, claims, citations, sources, or authorial attitudes that were not supplied.
+
+## 3. Preserve
+
+Compare source and revised text claim by claim. Verify:
+
+- numbers, ranges, dates, units, names, versions, and locations;
+- the direction and magnitude of each result;
+- who acted, who did not act, and who is responsible;
+- uncertainty, causation, scope, and conditions;
+- quotations, citation targets, and attribution;
+- legal, medical, safety, and scientific language the user requested or the artifact requires.
+
+Do not turn `not measured` into `zero`, a proposal into an outcome, a warning into a guarantee, or an estimate into an observed result. Preserve an explicit disclaimer when the user requests it or its removal would create a false impression. Keep one functional qualifier when it limits a claim.
+
+## 4. Re-check
+
+Run the detector on the newly written text, not on pre-existing file contents. Fix block hits. Read each warning in context and either revise it or keep it deliberately. Re-check the clean version. A zero count does not prove quality or human authorship; the fact comparison and read-aloud pass remain required.
+
+For files, review only the diff or newly generated section when the host can provide it. Exclude exact user quotations, code, internal plans/evidence/ledgers, and handoff/status records. For DOCX/PPTX, extract text immediately after creation and rerun the checks. PDF extraction must be supplied by the host adapter in Phase B.
+
+## Anti-overcorrection
+
+- Do not remove all qualifiers, all passive voice, all bullets, all adverbs, or all formal transitions.
+- Do not force contractions, first person, fragments, humor, or a casual register.
+- Do not flatten precise vocabulary just because it appears on a warning list.
+- Do not replace repeated field terminology with artificial synonyms.
+- Do not optimize an AI-detection score or promise that a detector will classify the text as human.
+- Do not erase a legal, medical, safety, or citation requirement to make prose shorter.
+
+Finish with two questions: does the new text still make exactly the same supported claims, and would this voice fit the writer and audience? If either answer is no, restore or revise the affected sentence.

package/skills/lit-humanizer/references/taxonomy.md

@@ -0,0 +1,153 @@
+# Pattern taxonomy
+
+These are editing cues, not authorship judgments. A block is high-confidence residue in a reader-facing artifact. A warning asks for contextual review; it is not a requirement to edit. Internal logs, ledgers, status output, legal notices, quotations, and deliberate author voice have different jobs. Machine rules live in rules.json; manual patterns are editorial checks.
+
+## 1. Deliverable labels, source footers, and limitation clutter
+
+This original category comes from the 26 real lines in fixtures/pos-real.txt and the adjacent research synthesis. Its central distinction is audience: internal evidence may need exact caveats, while a requested report should state the relevant limit once where it affects interpretation.
+
+### Category-1 rules
+
+- ko-bold-deliverable-label — ko, block. Bold process labels surface internal metadata. Rewrite the claim as a sentence.
+- en-bold-deliverable-label — en, block. Bold evidence labels create a detached mini-ledger.
+- ko-plain-meta-label and en-plain-meta-label — ko/en, block in prose. Connect a plain source or status line to its claim. A non-bold attribution directly below a Markdown table, image, figure or figcaption may remain, separated by no more than one blank line. A source line carrying a caveat still blocks. A source line standing alone in prose still blocks.
+- ko-stacked-limitation-chain and en-stacked-limitation-chain — ko/en, block for disclaimer chains. Keep each material boundary, but combine it into readable prose or place it where the reader needs it.
+- Workflow-vocabulary rules — en/ko, block. Internal stage names, evidence-bundle labels, tracking keys, and screenshot labels stay in evidence records. A required public citation remains a source citation.
+- en-demo-self-disclaimer — en, block. A demonstration label attached to a synthesis, assumption or recommendation is process framing. Name what is proposed and what was measured.
+- ko-bracket-status-tag and en-bracket-status-tag — ko/en, block. Replace bracket keys with a sentence describing the status.
+- ko-standalone-scope-disclaimer and en-standalone-scope-disclaimer — ko/en, block for stock disclaimers. Preserve a user-requested legal, medical or safety notice and place it once where it applies.
+- ko-status-negation, ko-mi-status, en-epistemic-negation and en-not-performed — warnings. A negative result or unperformed action may be the finding. Write the actor, scope and verb plainly; never turn unknown into false or an estimate into a fact.
+- ko-plain-meta-label-warn and en-plain-caveat-label — warnings for boundary headings. They may be useful in internal or technical packets.
+- en-knowledge-cutoff, en-ai-self-disclaimer, en-hope-helps and ko-chatbot-signoff — block. Remove the model wrapper or generic sign-off, then retain the answer.
+- ko-chatbot-opener and en-chatbot-opener — block stock greetings and praise at the start of a deliverable.
+- en-chatbot-closer — block the fixed offers to provide another version. en-closer-offer remains a warning for the broad “anything else” question: the negative corpus contains two ordinary task questions, so this wording alone cannot block.
+- en-lets-dive — block only the fixed dive-in and no-further-ado openers. A nearby invitation phrase remains outside the block rule because it appeared in a pre-2022 human manual.
+- en-testament: block the fixed significance formula. en-left-in-narration: block first-person editing narration left in an artifact.
+- code-ai-attribution — block generated-by trailers in commits and PRs under the family rule.
+
+### Real-line rewrite ledger
+
+Each row maps one of the 26 local positive lines to a usable sentence. It preserves the named source, value, scope, status, and uncertainty that make the line useful. Identifiers and process labels move to a citation or internal record when they are not reader-facing evidence.
+
+1. **Expected outcome in a source footer.**
+> Rewrite: “The Example Shipyard Sustainability Report 2024 lists 30% higher productivity and 30% shorter lead time as expected outcomes after smart-yard programme completion; neither is a realized result.”
+2. **Proposed control points in a source footer.**
+> Rewrite: “The control points are pilot demonstration recommendations drawn from the Example Shipyard yard digital twin and digital work-order descriptions.”
+3. **Architecture summary in a source footer.**
+> Rewrite: “The architecture summary draws on NVIDIA industrial digital-twin materials and describes a pilot demonstration synthesis.”
+4. **Assumptions in a source footer.**
+> Rewrite: “The proposed leading indicators and baselines are demonstration assumptions derived from Example Shipyard smart-yard programme expected outcomes.”
+5. **Claim and source IDs in a fact line.**
+> Rewrite: “Example Shipyard describes 30% higher productivity and 30% shorter lead time as expected after smart-yard programme completion, not as realized results.” Put the claim-to-source mapping in a footnote or internal record.
+6. **HTML evidence-boundary label.**
+> Rewrite: “This summary describes public NVIDIA product and technology pages; it does not establish deployment at a specific site.”
+7. **Table caption carrying a publisher record.**
+> Rewrite: “Table 2 separates evidence units by question. The supporting records are final report v3 and the publisher’s source record.” Keep an adjacent plain citation if the layout calls for one.
+8. **Global data mistaken for a persistence window.**
+> Rewrite: “The 100 ps global data do not establish persistence over 20–50 ps.”
+9. **Absence of positive evidence mistaken for proof of absence.**
+> Rewrite: “This upper bound does not positively establish the absence of transition or independent nucleation; the final direction remains undecided.”
+10. **Image caption with an internal receipt term.**
+> Rewrite: “Figure 6. Structure aligned to one 2.885 µs PyMOL frame. Source: the verified local frame and its image record.”
+11. **Unrun-work table row.**
+> Rewrite: “No remote check or new simulation ran; both require explicit approval.” Keep the full execution record in the internal ledger.
+12. **Non-observation versus physical absence.**
+> Rewrite: “The event was not observed; that does not show that the physical process was absent.”
+13. **Stable sensitivity set and unstarted simulation.**
+> Rewrite: “All nine sensitivity combinations were stable, but PMF MD has not started.”
+14. **Unvalidated model statement.**
+> Rewrite: “OPLS-AA has not been validated for epoxide; the report quantifies this limitation below.”
+15. **Fitting parameter statement.**
+> Rewrite: “We did not use a fitting parameter to match the experimental dissociation temperature.”
+16. **Mixed-system measurement statement.**
+> Rewrite: “This value was not measured directly in the mixed system.”
+17. **Classification-versus-size distinction.**
+> Rewrite: “This validates the present/absent classification, not the size estimate.”
+18. **Research and investment disclaimer chain.**
+> Rewrite: “연구 목적의 자료입니다. 투자 조언이 아니며, 실제 체결과 수익성은 확인하지 않았습니다. 미래 성과나 주문 체결도 보장하지 않습니다.” Preserve the legal boundary and each uncertainty.
+19. **Missing forecast.**
+> Rewrite: “SOXL 가격 전망은 이 자료에서 제공하지 않습니다.”
+20. **Source packet plus no-refetch caveat.**
+> Rewrite: “The source was a fact packet retrieved on 2026-07-13; the page was not fetched again or independently reproduced.”
+21. **Confidence label plus scope.**
+> Rewrite: “Confidence is high within the supplied fact packet. The claims were not independently revalidated.”
+22. **Normalized metric versus traded notional.**
+> Rewrite: “The value is normalized holdings change multiplied by the current close; it is not verified traded notional.”
+23. **Host verification, production run and mechanism claim.**
+> Rewrite: “The archive was not verified on the NVIDIA host, no production mdrun ran, and no mechanism closure is claimed.”
+24. **Uncertain image-sheet measurement.**
+> Rewrite: “높이는 130 mm입니다. 사진 시트에 적힌 표기이며 확정값이나 실측값은 아닙니다.”
+25. **Validation checklist phrased as meta text.**
+> Rewrite: “관련 검사와 실제 사용 경로를 확인하고, 근거와 가정을 구분합니다.”
+26. **Selection status plus unverified depth and permit result.**
+> Rewrite: “선별 결과에서 수심과 기준면은 아직 검증하지 않았습니다. 이 결과는 허가 판단이 아닙니다.”
+
+### Additional shapes from the synthesis
+
+- **Bracket status keys:** Replace a compact review tag with a sentence that states the condition.
+> Sample tag: [미검증] / [unverified]
+> Korean rewrite: 표 3의 추정치는 아직 확인하지 않았다.
+> English rewrite: The estimate in table 3 remains unverified.
+- **Evidence-status tags:** Replace a label with a sentence naming the source or measurement.
+> Sample tag: [증거 기반]
+> Rewrite: Name the source or measurement beside the claim.
+- **Mid-dot caveat chains:** Separate material conditions and state each only where it affects the claim.
+> Input: 미검증 · 허가 판단 아님 · 투자 조언 아님
+> Rewrite: 허가 여부는 관할 기관이 판단합니다. 이 분석은 투자 조언이 아닙니다.
+- **Standalone scope disclaimers:** Use a factual sentence only when the supporting facts are known. If authority or evidence is missing, name the responsible decision-maker or the specific gap.
+> Sample label: 허가 판단 결과 아님
+> Rewrite: The document summarizes the study; the permitting agency decides.
+- **Status negatives:** A negative status may be factual rather than stylistic. Name the action and scope; never imply completion.
+> Sample status: 미실행 / not performed
+> Rewrite: 원격 검증은 실행하지 않았습니다.
+- **Source footers carrying caveats:** Attach the source to a claim in a note, reference list or one plain caption line. A caveat on that line remains visible because it changes interpretation.
+- **Demonstration labels:** Replace a self-description with the actual demonstrator scope, or say which components are proposed. Keep whether a result is simulated or assumed.
+- **Workflow terms and tracking keys:** Replace internal stage and evidence labels with a reader-facing citation, or keep the exact record internally.
+- **The phrase “Evidence boundary”:** State the boundary once as a normal sentence beside the affected conclusion. Keep a formal section only when the technical or review format requires it.
+- **Bare “limitations” or “challenges” slots:** Name the specific open item, owner, test or date if known. Do not invent one to fill an empty slot.
+
+### Caption exemption contract
+
+A plain, non-bold attribution line directly following a Markdown table, image, HTML figure/caption, or figure/table caption may remain if it is attached or separated by no more than one blank line. A Korean materials attribution follows the same rule. No other prose inherits the exemption. A caveat on the line keeps it scannable. For example:
+> A source note that says the figures are forecasts rather than realized outcomes carries a caveat.
+
+Fixtures cover each attachment shape, one blank line, two blank lines, an embedded caveat, and an attribution line standing alone in prose.
+
+## 2. Structure and contrast
+
+- en-not-x-but-y — en, warn. Warn if the contrast opens a sentence or occurs at least twice in a document. A single natural mid-sentence contrast stays clean. Rewrite the positive claim only when the negative half corrects no real belief.
+- ko-negation-contrast — ko, warn. Review repeated negative-then-positive framing; keep a contrast that carries two distinct facts.
+- en-forced-triad and ko-forced-triad — en/ko, warn. Remove padded members, not a real three-part list.
+- en-colon-reveal — en, warn. A dramatic takeaway label can become a direct statement.
+- manual-en-cleft and manual-negative-list — editorial-only cues; no detector assertion is made.
+
+## 3. Openers and endings
+
+- en-throat-clearing and en-conclusion-opener — warn; remove only a redundant announcement.
+- ko-chatbot-invitation — warn; a service invitation may belong in chat but not in an independent report.
+- manual-fake-profound-kicker — manual review; end with a concrete finding, decision or next action if the original supports one.
+
+## 4. Attribution
+
+- en-weasel-attribution and ko-weasel-attribution — warn. Name an available source; never invent one to replace vague attribution.
+
+## 5. Hedge calibration
+
+- en-stacked-hedges and ko-stacked-hedges — warn. Keep the qualifier or qualifiers that set claim strength. Legal, medical, scientific, contractual and forecast language is not simplified to sound confident.
+
+## 6. Vocabulary
+
+- en-lexical-tell and ko-calque — warn. A word list prompts a meaning check; it is not a forbidden-word dictionary.
+- ko-double-passive and ko-light-verb — warn. Name an actor or a direct verb only when it preserves the source.
+
+## 7. Rhythm and formatting
+
+- en-em-dash-cluster and ko-decorative-bold — warn. Genre, source voice, accessibility and information hierarchy determine whether to edit.
+- Sentence length, paragraph length, list density, heading case and punctuation are manual-only unless an explicit machine rule exists.
+
+## 8. Code and developer writing
+
+- code-narrating-comment — warn. A comment should explain a constraint or reason.
+- code-defensive-guard — warn. Remove a branch only after verifying the type and trust boundary.
+- code-ai-attribution — block generated-model trailers in commit and PR text under the family rule.
+- manual-one-use-abstraction — editorial-only. Inline a one-use wrapper only when it adds no name, boundary, test seam or future reuse.