@litfamily/litopencode 1.0.7 → 1.0.10

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

@@ -0,0 +1,33 @@
+# Korean prose metrics
+
+The optional Node 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 canonical skill directory:
+
+~~~sh
+node scripts/ko-metrics.mjs --json path/to/draft.md
+node scripts/test-ko-metrics.mjs
+~~~
+
+The eight cases in fixtures/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.