@litfamily/lithermes 1.0.7 → 1.0.9

package/assets/lithermes-plugin/skills/lit-humanizer/references/en-patterns-structure.md

@@ -0,0 +1,375 @@
+# English patterns: structure, formatting, and source use
+
+These entries combine formatting and communication observations from HumanInk, humanizer, stop-slop, no-ai-slop, and Wikipedia’s field guide. A pattern is a reason to inspect a passage, not a reason to normalize every document.
+
+## EN-S01 | Negative-then-positive contrast
+- Overlap: HumanInk #9 (weight 6); humanizer 1; stop-slop binary contrast; Wikipedia negative parallelism.
+- What / why: A sentence creates an opponent or weak view, then announces the writer’s positive claim.
+- Detect: Look for paired negatives and positives; the current detector warns only for a sentence-opening contrast or two occurrences in a document.
+- Bad → better:
+> Bad: The issue is not the server, but the cache key.
+> Better: The cache key causes the issue.
+> Bad: This is not merely a form change; it is a new review process.
+> Better: The form change also adds a review step.
+- False positives: Keep the contrast when it corrects a claim the reader actually holds or when both halves add facts.
+- Evidence: weak; structure alone is not authorship evidence.
+
+## EN-S02 | Forced three-part list
+- Overlap: HumanInk #10 (5); humanizer 6; stop-slop rhythm check; Wikipedia rule-of-three category.
+- What / why: Related ideas are forced into a triad for a feeling of completeness.
+- Detect: Check whether all three members are distinct and needed.
+- Bad → better:
+> Bad: The meeting covered speed, quality, and excellence.
+> Better: The meeting covered speed and output quality.
+> Bad: The town needs innovation, inspiration, and investment.
+> Better: The town needs investment for the two planned repairs.
+- False positives: Keep three true alternatives, stages, or measurable dimensions.
+- Evidence: weak; lists of three are ordinary writing.
+
+## EN-S03 | Repeated sentence opening
+- Overlap: humanizer 7; stop-slop rhythm check.
+- What / why: Several neighboring sentences begin with the same subject or phrase without a rhetorical reason.
+- Detect: Compare sentence starts across a paragraph.
+- Bad → better:
+> Bad: The clerk checked the form. The clerk called the applicant. The clerk closed the case.
+> Better: The clerk checked the form, called the applicant, then closed the case.
+> Bad: It stores the key. It renews the key. It removes the key.
+> Better: It stores and renews the key, then removes it.
+- False positives: Keep an intentional refrain or legal sequence that relies on repeated wording.
+- Evidence: weak; repetition can be a deliberate voice choice.
+
+## EN-S04 | Staged question or false alternative
+- Overlap: humanizer 4–5; stop-slop rhetorical setup; HumanInk #21 (weight 9) for praise framing.
+- What / why: A question, confession, or rejected option announces an insight without adding support.
+- Detect: Ask whether the following sentence answers an objection that appears nowhere else.
+- Bad → better:
+> Bad: Have you ever wondered why the printer stops? The paper sensor is dirty.
+> Better: A dirty paper sensor stops the printer.
+> Bad: You might think the team needs another server, but the query is unindexed.
+> Better: The query lacks an index.
+- False positives: Keep a real interview question or a concern that the draft explicitly answers.
+- Evidence: weak; the exact setup can be effective in essays and conversation.
+
+## EN-S05 | Dramatic fragments and negative listing
+- Overlap: stop-slop structures; humanizer 2 and 5.
+- What / why: Short fragments build drama before stating one ordinary point.
+- Detect: Look for a row of fragments that can be stated as a complete claim.
+- Bad → better:
+> Bad: No handoff. No owner. No plan. Just a missed deadline.
+> Better: Nobody owned the handoff, so the deadline passed without a plan.
+> Bad: The button moved. Then the icon. Then the label.
+> Better: The redesign moved the button, icon, and label.
+- False positives: Preserve fragments in dialogue, poetry, or a voice that uses them intentionally.
+- Evidence: weak; cadence and genre matter.
+
+## EN-S06 | Colon reveal or announced takeaway
+- Overlap: humanizer 4; HumanInk #15 (6) for label lists; no-ai-slop colon check.
+- What / why: A short headline announces a routine fact as a reveal.
+- Detect: Inspect a colon whose text after it is one plain sentence.
+- Bad → better:
+> Bad: Key takeaway: the review took two days.
+> Better: The review took two days.
+> Bad: The surprise: both files were already present.
+> Better: Both files were already present.
+- False positives: Keep colons for lists, quotes, definitions, and real section labels.
+- Evidence: weak; a local editing cue.
+
+## EN-S07 | Transition-word chain
+- Overlap: HumanInk #27 (5); stop-slop transition guidance; Wikipedia signposting.
+- What / why: Repeated connectors announce every paragraph’s relation, even when the relation is obvious or absent.
+- Detect: Count repeated transitions in one section and check the logic each one expresses.
+- Bad → better:
+> Bad: Meanwhile, the clinic closed. On the other hand, the bus ran. That said, the road remained open.
+> Better: The clinic closed, but the bus kept running while the road stayed open.
+> Bad: Also, the launch moved. Moreover, the build took longer.
+> Better: The launch moved because the build took longer.
+- False positives: Keep a transition that clarifies a genuine contrast or consequence.
+- Evidence: weak; the source’s own threshold is a frequency check, not a block rule.
+
+## EN-S08 | Repeated one-line closer
+- Overlap: HumanInk #28 (7); humanizer 2 and 13; no-ai-slop ending review.
+- What / why: A final one-sentence paragraph repeats the point or adds a generic slogan.
+- Detect: Compare it with the immediately preceding paragraph and the opening.
+- Bad → better:
+> Bad: The cache cut repeat work. That is the real win.
+> Better: The cache cut repeat work.
+> Bad: The team shipped the fix Friday. The future is bright.
+> Better: The team shipped the fix Friday.
+- False positives: Keep a one-line ending that adds an action, date, or distinct conclusion.
+- Evidence: weak; a paragraph’s last line can carry deliberate emphasis.
+
+## EN-S09 | Uniform paragraph lengths
+- Overlap: HumanInk #29 (6); stop-slop rhythm checks.
+- What / why: Several paragraphs have the same sentence count and length despite changing subject.
+- Detect: A rough metric can flag four or more similarly sized prose paragraphs; review the content before editing.
+- Bad → better:
+> Bad: Four report sections each use three sentences of similar length.
+> Better: Let the one-sentence result stand alone and keep the methods section at its needed length.
+> Bad: Every chapter paragraph ends with a short summary line.
+> Better: Keep summary lines only where they help navigation.
+- False positives: Templates and manuals may need regular paragraph shapes.
+- Evidence: weak; uniformity is not proof and can aid scanning.
+
+## EN-S10 | Bold-label vertical list
+- Overlap: HumanInk #15 (6); humanizer 19; Wikipedia inline-header lists.
+- What / why: Each bullet repeats a bold category label and then explains it.
+- Detect: Scan for several consecutive bullets with the same label-and-colon structure.
+- Bad → better:
+> Bad: - Speed: The query is faster. - Safety: The gate is safer.
+> Better: The query is faster, and the gate rejects unsigned input.
+> Bad: - Owner: Mira reviews. - Date: Friday is due.
+> Better: Mira will finish the review by Friday.
+- False positives: Keep labels in dashboards, forms, slides, and reference lists.
+- Evidence: weak; formatting depends on the reader’s scanning task.
+
+## EN-S11 | Bullets replacing connected prose
+- Overlap: HumanInk #34 (5); humanizer 19; Wikipedia list-format observations.
+- What / why: A paragraph becomes a long list even though the items form one sentence or cause chain.
+- Detect: Ask whether items are independent and whether the list materially improves scanning.
+- Bad → better:
+> Bad: - The server restarted. - The queue cleared. - The alert stopped.
+> Better: The server restarted, which cleared the queue and stopped the alert.
+> Bad: - Rain began. - The field closed. - The match moved indoors.
+> Better: Rain forced the match indoors after the field closed.
+- False positives: Keep checklists, requirements, decisions, and true enumerations.
+- Evidence: weak; many professional genres rely on lists.
+
+## EN-S12 | Numbered steps for a simple sequence
+- Overlap: HumanInk #35 (5); humanizer layout categories.
+- What / why: A short procedure receives numbered packaging that implies complex stages.
+- Detect: Ask if sequence, prerequisites, or cross-reference require numbers.
+- Bad → better:
+> Bad: 1. Open the lid. 2. Pour the water. 3. Close the lid.
+> Better: Open the lid, pour the water, then close it.
+> Bad: Step 1: Save the file. Step 2: Open the folder.
+> Better: Save the file and open the folder.
+- False positives: Keep numbered steps when order is consequential or users need to cite a step.
+- Evidence: weak; procedure shape is task-dependent.
+
+## EN-S13 | Heading repeats title or body
+- Overlap: humanizer 20 and 24; Wikipedia heading and title categories.
+- What / why: A page title is repeated as an H1, or a heading is followed by a sentence that restates it.
+- Detect: Compare title, first heading, and first sentence for duplicate content.
+- Bad → better:
+> Bad: # Performance / Performance matters.
+> Better: # Performance / The page opens in 400 ms on the test device.
+> Bad: ## Results / The results were positive.
+> Better: ## Results / The error rate fell from 7% to 4%.
+- False positives: Some site templates require an H1 that matches the page title.
+- Evidence: weak; format and platform conventions decide.
+
+## EN-S14 | Title Case or heading-level decoration
+- Overlap: HumanInk #16 (4); Wikipedia title case, skipped levels, many top-level headings, and thematic breaks.
+- What / why: Headings capitalize every word, jump levels, or use repeated rules as decoration.
+- Detect: Follow the document or platform style guide; inspect hierarchy before changing case.
+- Bad → better:
+> Bad: ## Results From The Winter Trial
+> Better: ## Results from the winter trial
+> Bad: # Main / #### Detail / --------
+> Better: # Main / ## Detail
+- False positives: Keep proper names, publication styles, and headings required by an editor.
+- Evidence: hold as an authorship clue; it is ordinary style and platform variation.
+
+## EN-S15 | Bold used as decoration
+- Overlap: HumanInk #14 (4); humanizer 19; Wikipedia boldface category.
+- What / why: Many phrases are bolded, so emphasis loses contrast.
+- Detect: Review bold density in body text, not navigation labels.
+- Bad → better:
+> Bad: The **new** form **replaces** the **old** form **on Friday**.
+> Better: The new form replaces the old form on Friday.
+> Bad: The plan uses **three** separate **review** stages.
+> Better: The plan uses three separate review stages.
+- False positives: Keep emphasis that helps navigation or accessibility.
+- Evidence: hold as authorship evidence; the source severity is low and formatting varies by product.
+
+## EN-S16 | Emoji as professional formatting
+- Overlap: HumanInk #17 (5); Wikipedia emoji formatting.
+- What / why: Decorative symbols replace a clear heading or sentence.
+- Detect: Contextual review; do not remove icons that encode status consistently.
+- Bad → better:
+> Bad: 🚀 Launch: The service opens in Q3.
+> Better: The service opens in Q3.
+> Bad: ✅ Result: All 14 checks passed.
+> Better: All 14 checks passed.
+- False positives: Social posts, informal chat, and accessible status legends may use emoji deliberately.
+- Evidence: hold as a general signal; conventions differ.
+
+## EN-S17 | Em dash as a default connector
+- Overlap: HumanInk #13 (5); humanizer 8; Wikipedia em-dash category; stop-slop punctuation rule.
+- What / why: Dashes repeatedly bridge clauses without specifying their relationship.
+- Detect: Count clusters and compare with the writer’s sample and publication style.
+- Bad → better:
+> Bad: The build—after review—shipped—the same day—on Friday.
+> Better: The build shipped Friday after review.
+> Bad: The test—once restarted—passed—the second time.
+> Better: After a restart, the test passed on its second run.
+- False positives: A single dash can be natural; some writers use them as a signature.
+- Evidence: hold as a single-instance indicator; only a repeated cluster merits review.
+
+## EN-S18 | Hyphenated compound overuse
+- Overlap: humanizer 10; no-ai-slop hyphen check.
+- What / why: Compounds are hyphenated in positions where the house style uses open forms.
+- Detect: Compare attributive and predicative uses against the publication style guide.
+- Bad → better:
+> Bad: The report is high-quality and data-driven.
+> Better: The report is high quality and driven by the data.
+> Bad: The interface is user-friendly after launch.
+> Better: The interface is easy to use after launch.
+- False positives: Hyphens are often correct before a noun and in established terms.
+- Evidence: hold as authorship evidence; orthography varies.
+
+## EN-S19 | Parentheses and tables used as a holding bin
+- Overlap: Wikipedia unusual-table notes; humanizer format rules; stop-slop trust-the-reader principle.
+- What / why: Parentheses or dense tables collect details that the sentence or a note could state more clearly.
+- Detect: Ask whether each aside changes the main claim and whether a table has real comparison columns.
+- Bad → better:
+> Bad: The router (after restart) (following an alert) accepted the key.
+> Better: After the alert, the router accepted the key when it restarted.
+> Bad: A two-column table repeats each paragraph in shorter form.
+> Better: Keep the paragraph and use a table only for the actual comparison.
+- False positives: Preserve technical notes, citations, cross-reference tables, and parenthetical author voice.
+- Evidence: weak; layout purpose determines clarity.
+
+## EN-S20 | Broken or unexplained markup
+- Overlap: Wikipedia markup and template categories; humanizer 25; detector-research synthesis.
+- What / why: Leftover editor markup, malformed references, or tool tokens appear in prose.
+- Detect: Validate links, brackets, code spans, citation tokens, and rendering in the target host.
+- Bad → better:
+> Bad: The result{{ref?}} appears in [section 3.
+> Better: The result appears in section 3, with a valid citation.
+> Bad: The source marker is followed by an unexplained internal cite token.
+> Better: Replace it with the actual reference or remove it if no source exists.
+- False positives: Code examples and quoted markup are data, not prose defects.
+- Evidence: validated for a broken artifact that fails to render; it does not identify who wrote it.
+
+## EN-S21 | Citation integrity problems
+- Overlap: Wikipedia citation categories: broken links, invalid identifiers, unrelated references, incomplete book details, unused references, tracking parameters.
+- What / why: A citation may fail to support the claim even when its syntax looks plausible.
+- Detect: Resolve each link or identifier; check the cited page and the claim it supports.
+- Bad → better:
+> Bad: A report cites a dead URL with no archived copy.
+> Better: Replace the link with a working primary source, or state that support is unavailable.
+> Bad: A DOI opens a paper about a different method.
+> Better: Cite the paper that contains the stated method.
+- False positives: A broken link can have a human cause. Repair evidence instead of labeling authorship.
+- Evidence: validated as a reference-quality check, not a prose-origin signal.
+
+## EN-S22 | Placeholder text and pseudo-citations
+- Overlap: Wikipedia phrasal templates, placeholder text and reference-markup issues; HumanInk #20.
+- What / why: Draft notes or source-shaped placeholders remain where a real citation should be.
+- Detect: Search for empty names, template markers, or citations that do not resolve to an item.
+- Bad → better:
+> Bad: The population rose, according to [source needed].
+> Better: Add the source that measured the population or remove the unsupported claim.
+> Bad: The report cites an unnamed “recent survey.”
+> Better: Name the survey and its date if supplied; otherwise call the statement unverified.
+- False positives: A visible TODO is appropriate in a draft handed to an editor.
+- Evidence: weak for authorship, strong as a completion check.
+
+## EN-S23 | Process-heavy edit narration
+- Overlap: Wikipedia edit-summary categories; humanizer 25; local workflow-jargon cues.
+- What / why: The artifact narrates that it preserved, retained, or carefully complied instead of stating the revised content.
+- Detect: Ask whether the sentence describes the subject or the editing process.
+- Bad → better:
+> Bad: I preserved all relevant citations and fixed the formatting.
+> Better: The report cites the survey in note 4 and uses one heading style.
+> Bad: The change was made in order to retain the original behavior.
+> Better: The change keeps the original behavior.
+- False positives: Change logs, review notes, and audit records need process narration.
+- Evidence: weak in internal records; strong as leftover text in a standalone deliverable.
+
+## EN-S24 | Abrupt style shift
+- Overlap: Wikipedia pronounced-style-shift category; humanizer voice-preservation section.
+- What / why: One section differs sharply in tone, sentence structure, or vocabulary from its neighbors.
+- Detect: Compare adjacent sections and the author’s supplied samples; do not use a detector score.
+- Bad → better:
+> Bad: A plain report switches to a promotional paragraph midway through.
+> Better: Keep the stated facts and return to the report’s established register.
+> Bad: A first-person essay suddenly uses a formal generic biography voice.
+> Better: Preserve the author’s original point of view throughout.
+- False positives: A quoted passage, change in audience, or authored section boundary may explain the shift.
+- Evidence: weak; verify provenance and keep distinct voices intact.
+
+## EN-S25 | Internal citation or editor marker in published prose
+- Overlap: Wikipedia tool-specific citation residue and comment indicators; source-synthesis workflow keys.
+- What / why: A model or editing interface marker appears as if it were a reader-facing citation.
+- Detect: Review rendered output and link target, not just the text source.
+- Bad → better:
+> Bad: The sentence ends with an unexplained internal search-result token.
+> Better: Replace the token with a checked citation or remove it.
+> Bad: A comment includes a tool command as its only source.
+> Better: Link to the document or explain the source in ordinary prose.
+- False positives: An article explaining markup may quote these markers as examples.
+- Evidence: validated as publication cleanup; never infer authorship from the marker alone.
+
+## EN-S26 | Chat residue and reflexive agreement
+- Overlap: HumanInk #19 (weight 10), #21 (9), #33 (8); humanizer chatbot residue; no-ai-slop conversation residue.
+- What / why: A pasted greeting, praise, agreement, or invitation speaks to a chat user instead of advancing the document.
+- Detect: Check opening and final lines for service phrases, then remove them only when they are not part of a quoted exchange.
+- Bad → better:
+> Bad: Great question! The permit portal closes at 4 p.m.
+> Better: The permit portal closes at 4 p.m.
+> Bad: The new form has four sections. Let me know if you'd like me to explain them.
+> Better: The new form has four sections: contact, dates, eligibility, and signature.
+- False positives: Keep actual interview dialogue, user-facing service copy, or a requested conversational reply.
+- Evidence: validated for identifying pasted response residue; the phrase does not establish the origin of surrounding prose.
+
+## EN-S27 | Context-free modern-world opener
+- Overlap: HumanInk #25 (weight 8), #7 (9); Wikipedia generalization and significance categories.
+- What / why: A broad claim about the modern era delays the topic without giving a date, location, or change.
+- Detect: Search opening clauses for claims about today's world, changing conditions, or a new era; retain only a sourced time frame.
+- Bad → better:
+> Bad: In today's world, clinics need reliable appointment systems.
+> Better: Clinics need reliable appointment systems.
+> Bad: As we navigate an increasingly connected era, residents file forms online.
+> Better: Residents file the forms online.
+- False positives: Keep a dated historical comparison or a specifically sourced change over time.
+- Evidence: weak; the phrase family is a searchable editorial clue, not a universal marker.
+
+## EN-S28 | Passive voice
+- Overlap: stop-slop passive-voice structure; Wikipedia communication and source-attribution checks.
+- What / why: A passive can hide an actor who matters to the reader.
+- Detect: Ask whether the source names the actor and whether naming them improves the sentence. Do not infer an actor.
+- Bad → better:
+> Bad: The request was approved on Tuesday.
+> Better: The review board approved the request on Tuesday, if the source identifies it as the board's decision.
+> Bad: The notice was sent by the clinic on Monday.
+> Better: The clinic sent the notice on Monday.
+- False positives: Keep passive voice when the actor is unknown, irrelevant, protected, or intentionally omitted.
+- Evidence: weak; grammar is not an authorship signal. Attribution accuracy is the reason to review.
+
+## EN-S29 | Quote-style normalization
+- Overlap: HumanInk #18 (weight 3); document-specific typography conventions.
+- What / why: Straight and curly quotation marks are mixed without a house-style reason.
+- Detect: Compare the document with its style guide; do not scan quote glyphs as a universal defect.
+- Bad → better:
+> Bad: The notice calls the event “optional” and later says it is "required."
+> Better: The notice calls the event "optional" and later says it is "required." (when the guide specifies straight quotes)
+> Bad: The guide defines 'active status' as the current subscription state, then uses “active status” for the same term.
+> Better: The guide uses “active status” consistently for the subscription state. (when the guide specifies curly quotes)
+- False positives: Preserve exact source quotations when changing punctuation would violate a citation, transcript, legal, or archival requirement.
+- Evidence: hold; typography follows the publication's style guide and carries no authorship weight.
+
+## EN-S30 | False agency in objects and abstractions
+- Overlap: stop-slop false-agency structure; Wikipedia communication observations; humanizer active-voice guidance.
+- What / why: A document, dataset, rule, or abstract noun appears to decide or act when a person or system performed the action.
+- Detect: Find the verb's real actor in the supplied material. If the actor is unavailable, keep the original scope without inventing one.
+- Bad → better:
+> Bad: The dataset tells us that families prefer June.
+> Better: In the survey, 18 of 25 families chose June.
+> Bad: The complaint became a repair plan.
+> Better: The facilities team used the complaint to write a repair plan.
+- False positives: Conventional subjects such as “the report states” or “the system rejects” are clear when they name a real source or function.
+- Evidence: weak; check agency for clarity and attribution, not as a detector of authorship.
+
+## EN-S31 | Distant narrator and rhetorical setup
+- Overlap: stop-slop narrator-distance and sentence-starter checks; humanizer guidance on direct openings.
+- What / why: A sentence tells readers that an explanation is coming instead of stating it.
+- Detect: Remove the introductory question or announcement temporarily. Keep it only if it changes the reader's understanding.
+- Bad → better:
+> Bad: This is why the callback takes five seconds.
+> Better: The callback takes five seconds because the client reuses its token.
+> Bad: Nobody designed the paper form to be clear.
+> Better: The paper form puts the signature field below the fold.
+- False positives: Keep a genuine question, scene-setting voice, or attributed personal observation.
+- Evidence: weak; a direct start is often clearer, while narrative voice remains a valid choice.

package/assets/lithermes-plugin/skills/lit-humanizer/references/en-patterns.md

@@ -0,0 +1,57 @@
+# English pattern catalog
+
+Use this index to select the smallest reference shard that fits the draft. The shards merge overlapping pattern families from humanizer, HumanInk, no-ai-slop, stop-slop, and Wikipedia’s “Signs of AI writing” field guide. They are rewritten as editing guidance, not copied prompts.
+
+- Content, wording, claims: references/en-patterns-content.md
+- Structure, formatting, markup, source use, chat residue, passive voice, and quote style: references/en-patterns-structure.md
+- Editing rubric and checklist adapted from no-ai-slop eval and stop-slop: references/en-patterns-checklist.md
+- Code and developer prose: references/code-patterns.md
+- Detector rules: ../rules.json
+
+## HumanInk severity crosswalk
+
+Weights below preserve all 35 numeric severities from the pinned HumanInk catalog. A severity is a source-side editorial weight, not a calibrated authorship probability. When families merge, the shard names each contributing HumanInk ID and weight. Do not add these weights into a “probability” score.
+
+| ID | Weight | ID | Weight | ID | Weight | ID | Weight | ID | Weight |
+|---|---:|---|---:|---|---:|---|---:|---|---:|
+| 1 | 9 | 8 | 7 | 15 | 6 | 22 | 6 | 29 | 6 |
+| 2 | 7 | 9 | 6 | 16 | 4 | 23 | 7 | 30 | 7 |
+| 3 | 8 | 10 | 5 | 17 | 5 | 24 | 8 | 31 | 6 |
+| 4 | 8 | 11 | 6 | 18 | 3 | 25 | 8 | 32 | 5 |
+| 5 | 7 | 12 | 5 | 19 | 10 | 26 | 6 | 33 | 8 |
+| 6 | 6 | 13 | 5 | 20 | 8 | 27 | 5 | 34 | 5 |
+| 7 | 9 | 14 | 4 | 21 | 9 | 28 | 7 | 35 | 5 |
+
+## Overlap map
+
+| Combined family | Contributing source patterns |
+|---|---|
+| Significance, praise, promotional language, empty importance adjectives | HumanInk 1, 4, 32; humanizer 13, 16; no-ai-slop puffery; Wikipedia content categories on significance and promotion |
+| Borrowed authority and notability padding | HumanInk 2, 5; humanizer 17; Wikipedia coverage and vague attribution |
+| Shallow analysis attached to a fact | HumanInk 3; humanizer 15; stop-slop emphasis on unsupported interpretation |
+| Stock vocabulary, filler, and throat clearing | HumanInk 7, 22, 30; humanizer 4, 12; no-ai-slop word checklist; stop-slop phrase list |
+| Generic contrast, false range, or dead metaphor | HumanInk 9, 12, 26; humanizer 1, 3; stop-slop structures |
+| Repeated triads, transitions, and synonyms | HumanInk 10, 11, 27; humanizer 6, 7; stop-slop rhythm rules |
+| Weak source, cutoff, and speculative-gap language | HumanInk 20; humanizer 23; Wikipedia communication categories |
+| Generic conclusion or mirrored opening | HumanInk 6, 24, 28; humanizer 2, 13; no-ai-slop ending checks |
+| Decoration and list templates | HumanInk 14, 15, 16, 17, 18, 34, 35; humanizer 19–21; Wikipedia style categories |
+| Chat residue and sycophancy | HumanInk 19, 21, 33; humanizer 4, 22; no-ai-slop conversational residue check |
+| Paragraph and document rhythm | HumanInk 29, 31; humanizer 2, 6, 7; stop-slop rhythm rubric |
+| Announced points, emphasis, and broad modifiers | HumanInk 22, 30; no-ai-slop phrase checks; stop-slop filler and absolute-word checks |
+| Hidden actors and narrator distance | stop-slop passive, false-agency, and narrator structures; humanizer voice guidance; Wikipedia source attribution |
+
+## Wikipedia field-guide scope
+
+The linked field guide also discusses encyclopedic markup, citation integrity, edit summaries, and comment patterns. This port includes the parts that transfer to reports and developer writing: source verification, leftover interface markup, edit narration, formatting residue, and abrupt style shifts. Wikipedia-only rules about categories, templates, policy adherence, or article deletion are not applied as general prose rules.
+
+The field guide itself cautions that its indicators are context-dependent observations, not rules or proof of authorship. Read its “Ineffective indicators” and “Signs of human writing” material as false-positive safeguards. Older text, an em dash, a single word choice, or a clean detector result cannot establish who wrote a passage.
+
+The source page is [Wikipedia, “Signs of AI writing”](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing), an advice page maintained by WikiProject AI Cleanup. The topic headings are paraphrased in the catalog; examples are newly written. The upstream page was accessed 2026-09-25 and noted itself as needing updates on some current-model details.
+
+## Evidence labels
+
+- validated: an exact communication artifact or a source-reported comparison supports a narrow action, such as removing a pasted greeting or checking an invalid citation.
+- weak: the source supplies an editorial observation, severity weight, or heuristic, but not broad evidence that the form identifies generated text.
+- hold: preserve the form by default or require more context before editing; it is a style choice, convention, or known weak indicator.
+
+The evidence label is for editorial confidence, not authorship. All lexical lists are search aids. Every entry explains the clue, reason, detection hint, two bad-to-better pairs, and false-positive conditions.

package/assets/lithermes-plugin/skills/lit-humanizer/references/ko-metrics.md

@@ -0,0 +1,33 @@
+# Korean prose metrics
+
+The optional Python script turns a few source-tested patterns into repeatable counts. It is a review aid for a human editor. Counts do not measure authorship, writing quality, or whether a change is needed.
+
+## Signals and thresholds
+
+| Signal | Reported metric | Advisory threshold | Use |
+|---|---|---|---|
+| Conclusion-pivot clustering | Count of four conclusion or consequence pivots | 4 or more in one document | The source taxonomy treats counts above three as a reason to inspect a conclusion sequence. This is a frequency trigger, not a block rule. |
+| Korean cleft clustering | Count of common “what matters/what is needed” sentence shapes | 2 or more in at least 40 whitespace-delimited units | The source reports a corpus difference for cleft-like forms. The 40-unit floor is a local guard against flagging a tiny excerpt; the threshold has not been calibrated as a classifier. |
+| Paragraph-initial connectives | Number of listed connectives at sentence starts in each paragraph | 3 or more in one paragraph | The source discusses three repeated openers as an inspection point and reports model-dependent results. Preserve connectors that clarify logic. |
+| Connective-ending commas | Count and share of listed Korean connective endings followed by a comma | At least 3 comma hits and a share of 0.5 or more | This is a surface approximation. A comma may mark a real clause boundary; read the sentence before editing. |
+
+The JSON also reports sentence count, comma-bearing sentence share, sentence-length mean and variation, ending variety, and a balance-phrase count. These remain descriptive; this port sets no warning thresholds for them. Number-grouping commas are excluded from the prose comma count.
+
+The pivot set is 결론적으로, 따라서, 이를 통해, 그러므로. The paragraph-opening set is 또한, 따라서, 즉, 나아가, 아울러, 게다가, 더욱이, 그러므로, 반면, 그러나, 하지만. The connective-ending approximation counts 고, 며, 지만, 면서, 아서, 어서 when a comma follows. The cleft counter checks common 필요/중요/핵심/문제/관건/답 constructions.
+
+## Method limits
+
+The script removes Markdown headings, tables, blockquotes, inline code, and fenced code before counting. Sentence boundaries and whitespace-delimited Korean units are approximations; the script uses no morphological parser. Slides, legal text, notices, dialogue, procedures, and quotations can naturally cluster these forms. A zero result does not mean the prose is natural, and a warning does not call for a rewrite by itself.
+
+Do not combine metric values into a score or use them to rank writers or drafts. Compare a signal with the document's genre, source, audience, and purpose; preserve facts, scope, register, and purposeful repetition.
+
+## Run and fixtures
+
+From the installed plugin root:
+
+~~~sh
+python3 skills/lit-humanizer/scripts/ko_metrics.py --json path/to/draft.md
+LITHERMES_PYTHON=... npm --prefix packages/lithermes-installer run test:python
+~~~
+
+The eight cases in `examples/detector-validation/ko-metrics-golden.json` include a positive and a clean comparison for each threshold family. They exercise counts and minimum-length gates, not authorship labels. Extend them with paired examples before changing a threshold.