start-vibing 4.4.3 → 4.4.4

package/template/.claude/skills/research/references/research-methodology.md

@@ -0,0 +1,794 @@
+# Research Methodology — Reference
+
+> Read this when running the `research` skill. Every section is prescriptive: pick the framework that fits the question, cite primary sources, prefer triangulation over volume, and treat freshness as a per-claim property.
+
+Evidence protocol used throughout the skill: **URL + QUOTE + ACCESSED-AT + VERIFY-METHOD**. Every claim in `/docs/research/<topic>.md` must carry all four.
+
+---
+
+## 1. Academic Frameworks
+
+### 1.1 PRISMA 2020
+
+**PRISMA 2020** (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) is the canonical reporting standard for systematic reviews.
+
+- Page MJ, McKenzie JE, Bossuyt PM, et al. "The PRISMA 2020 statement: an updated guideline for reporting systematic reviews." _BMJ_ 2021;372:n71. doi:10.1136/bmj.n71. <https://www.bmj.com/content/372/bmj.n71>
+- Page MJ, Moher D, Bossuyt PM, et al. "PRISMA 2020 explanation and elaboration." _BMJ_ 2021;372:n160. doi:10.1136/bmj.n160.
+
+**27-item checklist** spans: title, abstract, introduction (rationale, objectives), methods (eligibility, sources, search, selection, extraction, RoB, synthesis), results, discussion, other (registration, support).
+
+**4-phase flow diagram** (Identification → Screening → Eligibility → Included) records counts at each step. Required for any systematic review claiming PRISMA conformance.
+
+**Extensions** (use whichever matches the review type):
+
+| Extension     | Use case                    | Citation                                                                        |
+| ------------- | --------------------------- | ------------------------------------------------------------------------------- |
+| PRISMA-ScR    | Scoping reviews             | Tricco AC et al. _Ann Intern Med_ 2018;169(7):467. doi:10.7326/M18-0850         |
+| PRISMA-S      | Search reporting            | Rethlefsen ML et al. _Syst Rev_ 2021;10:39. doi:10.1186/s13643-020-01542-z      |
+| PRISMA-P      | Protocols                   | Moher D et al. _Syst Rev_ 2015;4:1. doi:10.1186/2046-4053-4-1                   |
+| PRISMA-DTA    | Diagnostic test accuracy    | McInnes MDF et al. _JAMA_ 2018;319(4):388. doi:10.1001/jama.2017.19163          |
+| PRISMA-IPD    | Individual participant data | Stewart LA et al. _JAMA_ 2015;313(16):1657. doi:10.1001/jama.2015.3656          |
+| PRISMA-Equity | Health equity               | Welch V et al. _PLoS Med_ 2012;9(10):e1001333. doi:10.1371/journal.pmed.1001333 |
+| PRISMA-Harms  | Adverse events              | Zorzela L et al. _BMJ_ 2016;352:i157. doi:10.1136/bmj.i157                      |
+
+### 1.2 Cochrane Handbook v6.x
+
+**Cochrane Handbook for Systematic Reviews of Interventions, Version 6.4** (Higgins JPT, Thomas J, Chandler J, Cumpston M, Li T, Page MJ, Welch VA, eds). Cochrane, 2023. <https://training.cochrane.org/handbook>. The reference standard for intervention reviews.
+
+Mandates **PICO** framing: Population, Intervention, Comparator, Outcome (Richardson WS et al. _ACP J Club_ 1995;123:A12).
+
+Risk-of-bias tools:
+
+- **RoB 2** (randomized trials): Sterne JAC et al. "RoB 2: a revised tool for assessing risk of bias in randomised trials." _BMJ_ 2019;366:l4898. doi:10.1136/bmj.l4898. Five domains: randomization, deviations, missing data, measurement, reporting.
+- **ROBINS-I** (non-randomized): Sterne JAC et al. _BMJ_ 2016;355:i4919. doi:10.1136/bmj.i4919. Seven domains including confounding and selection.
+
+### 1.3 Scoping Reviews
+
+When the question is "what is out there" rather than "does X work":
+
+- **Arksey H, O'Malley L.** "Scoping studies: towards a methodological framework." _Int J Soc Res Methodol_ 2005;8(1):19–32. doi:10.1080/1364557032000119616. The original 5-stage framework.
+- **Levac D, Colquhoun H, O'Brien KK.** "Scoping studies: advancing the methodology." _Implement Sci_ 2010;5:69. doi:10.1186/1748-5908-5-69. Refines stage 4 (charting) and stage 5 (consultation).
+- **Pham MT, Rajić A, Greig JD, Sargeant JM, Papadopoulos A, McEwen SA.** "A scoping review of scoping reviews." _Res Synth Methods_ 2014;5(4):371–385. doi:10.1002/jrsm.1123.
+- **Peters MDJ, Marnie C, Tricco AC, et al.** "Updated methodological guidance for the conduct of scoping reviews." _JBI Evid Synth_ 2020;18(10):2119–2126. doi:10.11124/JBIES-20-00167. Current JBI standard.
+
+### 1.4 Rapid Reviews
+
+When timeline is weeks, not months:
+
+- **Tricco AC, Antony J, Zarin W, et al.** "A scoping review of rapid review methods." _BMC Med_ 2015;13:224. doi:10.1186/s12916-015-0465-6.
+- **Haby MM, Chapman E, Clark R, Barreto J, Reveiz L, Lavis JN.** "What are the best methodologies for rapid reviews of the research evidence for evidence-informed decision making in health policy and practice?" _Health Res Policy Syst_ 2016;14:83. doi:10.1186/s12961-016-0155-7.
+
+Standard concessions: single screener, restricted database set (e.g., MEDLINE only), no manual citation chasing, simplified RoB.
+
+### 1.5 Umbrella Reviews
+
+Reviews-of-reviews:
+
+- **Aromataris E, Fernandez R, Godfrey CM, Holly C, Khalil H, Tungpunkom P.** "Summarizing systematic reviews: methodological development, conduct and reporting of an umbrella review approach." _Int J Evid Based Healthc_ 2015;13(3):132–140. doi:10.1097/XEB.0000000000000055.
+
+### 1.6 SALSA Typology and Decision Tree
+
+**Grant MJ, Booth A.** "A typology of reviews: an analysis of 14 review types and associated methodologies." _Health Info Libr J_ 2009;26(2):91–108. doi:10.1111/j.1471-1842.2009.00848.x. The SALSA acronym (Search, AppraisaL, Synthesis, Analysis) lets you compare review types on shared dimensions.
+
+**Decision tree** for the research skill:
+
+```
+Is the question "does X work / what is the effect size"?
+  ├─ YES, with appraisal capacity (weeks–months)  → Systematic review (Cochrane + PRISMA 2020)
+  └─ NO
+     │
+     Is the question "what literature exists / what concepts"?
+       ├─ YES → Scoping review (PRISMA-ScR + JBI 2020)
+       └─ NO
+          │
+          Is the question "summarize prior reviews"?
+            ├─ YES → Umbrella review (Aromataris 2015)
+            └─ NO
+               │
+               Is the deadline < 4 weeks?
+                 ├─ YES → Rapid review (Tricco 2015)
+                 └─ NO  → Narrative review (default for skill output)
+```
+
+The research skill defaults to **scoping-style narrative output** with PRISMA-ScR-inspired transparency (sources searched, dates, inclusion criteria).
+
+### 1.7 EQUATOR Network and Reporting Standards
+
+**EQUATOR Network** <https://www.equator-network.org/> hosts 500+ reporting guidelines. For each study type cited in research output, prefer studies that conform: CONSORT (RCTs), STROBE (observational), CARE (case reports), SRQR (qualitative), AGREE II (guidelines).
+
+### 1.8 Gray Literature
+
+Theses, government reports, preprints, conference proceedings, working papers — often where the freshest evidence lives.
+
+- **Paez A.** "Gray literature: An important resource in systematic reviews." _J Evid Based Med_ 2017;10(3):233–240. doi:10.1111/jebm.12266.
+- **Adams J, Hillier-Brown FC, Moore HJ, et al.** "Searching and synthesising 'grey literature' and 'grey information' in public health: critical reflections on three case studies." _Syst Rev_ 2016;5:164. doi:10.1186/s13643-016-0337-y.
+
+Sources: OpenGrey (defunct as of 2020 but archived), GreyNet, Google Scholar (filtered to non-journal results), institutional repositories, ProQuest Dissertations, OSF preprints, SSRN, arXiv, bioRxiv, medRxiv, government .gov / .gov.uk / Europa portals.
+
+---
+
+## 2. Industry Methodologies (the Baymard Benchmark)
+
+### 2.1 Baymard Institute
+
+The gold standard for e-commerce UX research and the benchmark this skill emulates for source-first rigor.
+
+- **Baymard Institute methodology**: 700+ guidelines, 250+ benchmarked sites, 130,000+ hours of research as of 2024 (per <https://baymard.com/about/methodology>). Severity scoring 1–5 per UX issue, evidence-anchored quotes from think-aloud sessions, paired with screenshots and video. _Adapt this rigor: every research finding must carry an extracted QUOTE and a verifiable URL._
+- **Holst CC** ("Baymard's lead researcher") publishes per-issue tear-downs that combine quantitative benchmark scores with qualitative session evidence. The composite is the differentiator.
+
+### 2.2 Nielsen Norman Group
+
+- **Nielsen J.** "10 Usability Heuristics for User Interface Design." 1994; revised 2024. <https://www.nngroup.com/articles/ten-usability-heuristics/>. Visibility of system status; match between system and real world; user control; consistency; error prevention; recognition over recall; flexibility; minimalist design; help users recognize/diagnose/recover from errors; help and documentation.
+- **Nielsen J.** "How to Conduct a Heuristic Evaluation." 1994. <https://www.nngroup.com/articles/how-to-conduct-a-heuristic-evaluation/>. Three-to-five evaluators, independent passes, consolidated severity ratings.
+- **Nielsen J, Landauer TK.** "A mathematical model of the finding of usability problems." _INTERCHI '93 Proceedings_ 1993:206–213. doi:10.1145/169059.169166. The "5 users find 85% of issues" result — interpret as "5 users in _one_ round; iterate".
+
+### 2.3 Gartner
+
+- **Magic Quadrant**: 2-axis evaluation (Ability to Execute × Completeness of Vision). Methodology: <https://www.gartner.com/en/research/methodologies/magic-quadrants-research>. Cite the report ID and publication date — quadrant positions shift annually.
+- **Hype Cycle**: Innovation Trigger → Peak of Inflated Expectations → Trough of Disillusionment → Slope of Enlightenment → Plateau of Productivity. Fenn J, Raskino M. _Mastering the Hype Cycle_. Harvard Business Press, 2008. ISBN 978-1422121108.
+- **Peer Insights** is end-user reviewed; quantity-weighted, not authority-weighted — treat as supporting evidence, not primary.
+
+### 2.4 Forrester Wave
+
+Scoring across Current Offering, Strategy, Market Presence — each with weighted sub-criteria published per report. Methodology: <https://www.forrester.com/policies/forrester-wave-methodology/>. Vendors are categorized Leaders / Strong Performers / Contenders / Challengers.
+
+### 2.5 IDEO Method Cards
+
+51 cards across **Learn / Look / Ask / Try**. IDEO. _Method Cards: 51 Ways to Inspire Design._ William Stout, 2003. Use as a sampling menu for qualitative research design (e.g., shadowing, fly-on-the-wall, error analysis, role-playing).
+
+### 2.6 McKinsey 7-Step Problem Solving
+
+1. Define the problem
+2. Structure (issue tree, MECE)
+3. Prioritize
+4. Plan analysis & work
+5. Conduct analysis
+6. Synthesize findings
+7. Communicate (Pyramid Principle)
+
+Sources: Rasiel EM. _The McKinsey Way_. McGraw-Hill, 1999. ISBN 978-0070534483. Minto B. _The Pyramid Principle_. Pearson, 1987. ISBN 978-0273710516.
+
+**MECE** = Mutually Exclusive, Collectively Exhaustive — used to validate issue trees. **Hypothesis trees** drive top-down analysis: state the hypothesis, list the supporting sub-hypotheses, gather data to confirm/refute each leaf.
+
+### 2.7 BCG / Bain Frameworks
+
+- **5C** (Customer, Company, Competitors, Collaborators, Context)
+- **Porter 5 Forces**: Porter ME. _Competitive Strategy_. Free Press, 1980. ISBN 978-0029253601. Threat of new entrants, supplier power, buyer power, threat of substitutes, rivalry.
+- **VRIO**: Barney JB. "Firm resources and sustained competitive advantage." _J Manage_ 1991;17(1):99–120. doi:10.1177/014920639101700108. Valuable, Rare, Inimitable, Organized to capture value.
+
+### 2.8 Industry vs Academic Tradeoffs
+
+| Dimension       | Industry                 | Academic                        |
+| --------------- | ------------------------ | ------------------------------- |
+| Speed           | Days–weeks               | Months–years                    |
+| Reproducibility | Often proprietary        | Required (PRISMA, registration) |
+| Sample          | Convenience, paid panels | Probability where possible      |
+| Bias disclosure | Sponsor often hidden     | COI mandated                    |
+| Citation depth  | Selective                | Exhaustive                      |
+| Output          | Decision-ready           | Knowledge-ready                 |
+
+Research output for the skill blends both: industry speed + academic transparency (every claim cited, methods section in every doc).
+
+---
+
+## 3. Knowledge Ontology
+
+### 3.1 Formal Stack
+
+- **OWL 2** (Web Ontology Language). W3C Recommendation, 2012. <https://www.w3.org/TR/owl2-overview/>. DL/EL/QL/RL/Full profiles. Class expressions, property characteristics, axioms.
+- **RDF 1.1**. W3C Recommendation, 2014. <https://www.w3.org/TR/rdf11-concepts/>. Subject-predicate-object triples, IRIs, literals, blank nodes.
+- **RDFS** (RDF Schema). <https://www.w3.org/TR/rdf-schema/>. Lightweight class/property hierarchy on top of RDF.
+- **SKOS** (Simple Knowledge Organization System). W3C Recommendation, 2009. <https://www.w3.org/TR/skos-reference/>. Concept schemes for thesauri/taxonomies — _the right level of formality for `/docs/research/`_.
+
+### 3.2 Topic Maps vs RDF
+
+- **ISO/IEC 13250** Topic Maps. Topics, associations, occurrences. More navigation-oriented than RDF; weaker tool ecosystem after 2010. Pepper S. "The TAO of Topic Maps." 2002. <https://ontopia.net/topicmaps/materials/tao.html>.
+- **RDF/SKOS** wins on tooling (Protégé, Apache Jena, GraphDB), serializations (Turtle, JSON-LD), and ecosystem (Wikidata, schema.org).
+
+### 3.3 Upper Ontologies
+
+- **BFO 2.0** (Basic Formal Ontology). Arp R, Smith B, Spear AD. _Building Ontologies with Basic Formal Ontology._ MIT Press, 2015. ISO/IEC 21838-2:2021. Continuant/occurrent split.
+- **DOLCE** (Descriptive Ontology for Linguistic and Cognitive Engineering). Gangemi A et al. EKAW 2002. doi:10.1007/3-540-45810-7_18.
+- **SUMO** (Suggested Upper Merged Ontology). Pease A. <http://www.adampease.org/OP/>.
+- **Cyc**. Lenat DB. "CYC: A large-scale investment in knowledge infrastructure." _Commun ACM_ 1995;38(11):33–38. doi:10.1145/219717.219745.
+- **UFO** (Unified Foundational Ontology). Guizzardi G. _Ontological Foundations for Structural Conceptual Models_. PhD thesis, Twente, 2005.
+
+For LLM-consumable docs, _do not_ import upper ontologies — too heavy. Borrow concepts (continuant/occurrent for "things that persist vs events").
+
+### 3.4 Domain Vocabularies
+
+- **FOAF** (Friend of a Friend). <http://xmlns.com/foaf/spec/>.
+- **Dublin Core**. ISO 15836. <https://www.dublincore.org/specifications/dublin-core/dcmi-terms/>. 15 core elements: title, creator, subject, description, publisher, contributor, date, type, format, identifier, source, language, relation, coverage, rights.
+- **schema.org** (Google/Bing/Yahoo/Yandex). <https://schema.org/>. JSON-LD vocabulary for web data — used in research output frontmatter.
+- **SNOMED CT** (medicine). <https://www.snomed.org/>.
+- **Gene Ontology**. Ashburner M et al. _Nat Genet_ 2000;25:25–29. doi:10.1038/75556. <http://geneontology.org/>.
+
+### 3.5 Distinctions
+
+| Term                  | Structure         | Relations                          | Example                 |
+| --------------------- | ----------------- | ---------------------------------- | ----------------------- |
+| Controlled vocabulary | Flat list         | None                               | List of approved tags   |
+| Folksonomy            | Flat              | Implicit (co-occurrence)           | Twitter hashtags        |
+| Taxonomy              | Tree              | is-a only                          | Linnaean classification |
+| Thesaurus             | Tree + cross-refs | broader/narrower/related, synonyms | LCSH, MeSH              |
+| Ontology              | Graph             | Arbitrary typed relations + axioms | OWL ontology            |
+| Knowledge graph       | Graph + instances | Typed, with data                   | Wikidata                |
+
+ISO 25964-1:2011 / 25964-2:2013 specifies thesaurus interoperability — the SKOS-shaped sweet spot.
+
+### 3.6 Tooling
+
+- **Protégé** (Stanford). <https://protege.stanford.edu/>. The reference OWL editor.
+- **Apache Jena**. <https://jena.apache.org/>. Java RDF/SPARQL toolkit.
+- **rdflib** (Python). <https://rdflib.readthedocs.io/>.
+- **Wikidata SPARQL**. <https://query.wikidata.org/>. Live triples for entity verification.
+
+### 3.7 Relationship Types Beyond `is-a` / `has-a`
+
+- **Mereology** (parthood). Simons P. _Parts: A Study in Ontology_. OUP, 1987. ISBN 978-0198249436. `part-of`, `proper-part-of`, `overlaps`.
+- **Allen interval algebra** (temporal). Allen JF. "Maintaining knowledge about temporal intervals." _Commun ACM_ 1983;26(11):832–843. doi:10.1145/182.358434. 13 basic relations: before, meets, overlaps, starts, during, finishes, equals (plus inverses).
+- **OWL property characteristics**: `equivalent-to`, `disjoint-from`, `inverse-of`, `transitive`, `symmetric`, `reflexive`, `functional`, `inverse-functional`.
+- **SKOS semantic relations**: `broader`, `narrower`, `related`; lexical: `prefLabel`, `altLabel`, `hiddenLabel`; documentation: `definition`, `scopeNote`, `example`.
+
+### 3.8 Knowledge Graphs
+
+- **Wikidata**. Vrandečić D, Krötzsch M. "Wikidata: a free collaborative knowledgebase." _Commun ACM_ 2014;57(10):78–85. doi:10.1145/2629489.
+- **DBpedia**. Auer S et al. ISWC 2007. doi:10.1007/978-3-540-76298-0_52.
+- **YAGO**. Suchanek FM, Kasneci G, Weikum G. WWW 2007. doi:10.1145/1242572.1242667.
+
+### 3.9 Practical Lightweight Ontology for LLMs
+
+LLMs do not benefit from formal OWL reasoning. They benefit from:
+
+1. **Stable concept slugs** (`react-server-components`, not "RSCs").
+2. **Closed vocabulary of relations** (≤ 13 verbs — see `ontology-patterns.md`).
+3. **Inline encoding in markdown** (`Button is-a InteractiveElement`) — recoverable by `grep`.
+4. **Index files** (`/docs/research/index.md`) listing every concept and its file.
+5. **Backlinks** so a concept's full neighborhood is one query away.
+
+The skill's `extract-claims.py` walks the markdown and emits a JSON-LD slice — the round-trip rule is that any relation that fails extraction is mis-encoded.
+
+---
+
+## 4. Triangulation
+
+### 4.1 Denzin's Four Types
+
+**Denzin NK.** _The Research Act: A Theoretical Introduction to Sociological Methods_. Aldine, 1978 (3rd ed. McGraw-Hill, 1989). ISBN 978-0070164802.
+
+| Type                             | Definition                                  | Application in skill                                          |
+| -------------------------------- | ------------------------------------------- | ------------------------------------------------------------- |
+| **Data triangulation**           | Multiple data sources (time, space, person) | 3+ independent URLs per claim                                 |
+| **Investigator triangulation**   | Multiple researchers                        | Cross-check between research-query and research-verify agents |
+| **Theoretical triangulation**    | Multiple theoretical lenses                 | Compare framework views (e.g., NN/g + Baymard + ISO 9241)     |
+| **Methodological triangulation** | Multiple methods                            | Doc + code + screenshot evidence                              |
+
+### 4.2 Why "3 Sources" Fails
+
+Three articles can all cite the same upstream source. The classic case:
+
+- **Citogenesis** (term coined by Randall Munroe, _xkcd_ #978, 2011, <https://xkcd.com/978/>): Wikipedia statement → news article cites Wikipedia → Wikipedia citation updated to news article → fact is now "verified".
+- **Ward Cunningham**, inventor of the wiki, repeatedly warned about this loop in the WikiWikiWeb era; see <http://wiki.c2.com/?CitationNeeded>.
+
+**Mitigation**: trace each source to a _primary_ artifact (peer-reviewed paper, official spec, vendor docs at canonical URL, court filing, public dataset). Republished blog posts do not increase confidence.
+
+### 4.3 Source Independence
+
+Independence checks before counting a source:
+
+1. **Ownership tree**: who owns the publisher? (e.g., G/O Media properties share editorial pipelines).
+2. **Republication detection**: search a distinctive 8–12-word string from the article in quotes — if it appears across 5+ sites with different domains, it is syndicated.
+3. **Citation cascade**: if every recent article cites one 2019 paper, the 2019 paper is the source — not the citing articles.
+4. **Author overlap**: same byline across "independent" outlets means one informant.
+5. **Funding overlap**: same funder/sponsor across sources is _not_ independence.
+
+### 4.4 Cross-Disciplinary Triangulation
+
+A claim from one field gains confidence when an adjacent field arrives at it independently. Example: "code review reduces defects" — software engineering (Fagan ML. _IBM Syst J_ 1976;15(3):182–211. doi:10.1147/sj.153.0182) and behavioral economics (peer accountability literature) converge.
+
+### 4.5 Temporal Triangulation
+
+The same observation across multiple time points reduces single-period bias. For software/web: check a claim at t-12mo, t-6mo, t-now via Wayback Machine.
+
+### 4.6 Confidence Calibration Off Triangulation
+
+| Independent primary sources | Same-direction agreement | Confidence                        |
+| --------------------------- | ------------------------ | --------------------------------- |
+| ≥ 3                         | All agree                | High                              |
+| 2                           | All agree                | Medium                            |
+| 1 + corroborating secondary | Agree                    | Medium-low                        |
+| 1                           | n/a                      | Low (flag as single-source)       |
+| ≥ 2                         | Disagreement             | Contested — record both, no claim |
+
+### 4.7 Red Flags
+
+- **AI content farms**: clean prose, balanced bullets, no author photo, generic byline ("staff writer"), publication date without "updated", no DOI, no direct quotes from interviewees. See _Futurism_ exposés on AI-generated articles at _Sports Illustrated_ and _CNET_ (2023).
+- **Press-release regurgitation**: identical paragraphs across 6+ outlets within 24h.
+- **SEO copycats**: H2s match exactly across the top 10 search results.
+- **Pop-up "research firms"** with no addressable office and no methodology page.
+
+---
+
+## 5. Query Engineering
+
+### 5.1 Boolean and Operators
+
+`AND`, `OR`, `NOT` (or `-` in many engines). Most engines short-circuit `OR`-heavy queries; nest with parentheses. Example MEDLINE-style: `(("react"[Title] OR "next.js") AND ("data fetching"[MeSH]) NOT "react native")`.
+
+### 5.2 Proximity
+
+`NEAR/n` (varies by engine), `"phrase"` (exact). Google's `AROUND(n)` is undocumented but functional: `"server components" AROUND(5) "streaming"`.
+
+### 5.3 Truncation / Wildcards
+
+`*` matches one or more words (Google) or one or more characters (most academic DBs). `?` for single char in some DBs.
+
+### 5.4 Phrase / Field
+
+- `"exact phrase"` for verbatim
+- Field operators (Google): `intitle:`, `inurl:`, `intext:`, `site:`, `filetype:`, `before:YYYY-MM-DD`, `after:YYYY-MM-DD`, `cache:`
+- Academic field tags: PubMed `[Title]`, `[MeSH]`, `[au]`; Scopus `TITLE-ABS-KEY()`; WoS `TS=()`.
+
+### 5.5 Controlled Vocabularies
+
+- **MeSH** (Medical Subject Headings). NLM. <https://www.ncbi.nlm.nih.gov/mesh>. Hierarchical, manually curated. Use `[MeSH]` field tag in PubMed.
+- **ACM CCS** 2012 (Computing Classification System). <https://dl.acm.org/ccs>. ~2,000 concepts, computer science.
+- **IEEE Thesaurus**. <https://www.ieee.org/publications/services/thesaurus.html>.
+- **LCSH** (Library of Congress Subject Headings). <https://id.loc.gov/authorities/subjects.html>.
+
+### 5.6 Semantic vs Keyword
+
+Semantic search (Semantic Scholar, Elicit, Consensus) embeds query and corpus into a vector space; matches on meaning. Keyword search matches lexically. Use semantic for concept retrieval, keyword for exact-match verification of quotes.
+
+### 5.7 Query Decomposition
+
+Break a complex question into orthogonal sub-queries:
+
+```
+Q: "What's the state of React server components in late 2025?"
+  ├─ "react server components" specification 2024..2026
+  ├─ "react server components" production case studies
+  ├─ "react server components" performance benchmarks 2025
+  ├─ "react server components" criticism limitations
+  └─ "react server components" alternatives (qwik, solid, astro)
+```
+
+### 5.8 Bridge Queries
+
+When a query returns nothing, walk one step out:
+
+- Replace specific term with category (`"vite-plugin-foo"` → `vite plugin <feature>`)
+- Replace year filter with range (`2025` → `2024..2026`)
+- Drop domain filter (remove `site:example.com`)
+- Translate via known synonym (use the controlled-vocab thesaurus)
+
+### 5.9 Snowballing
+
+- **Backward snowballing**: read paper P, follow its citations.
+- **Forward snowballing**: who cites P? Use Google Scholar's "Cited by" or Semantic Scholar's citation graph.
+- **Wohlin C.** "Guidelines for snowballing in systematic literature studies and a replication in software engineering." _EASE '14_ 2014. doi:10.1145/2601248.2601268.
+
+### 5.10 Time-Boxing and Negative-Space Queries
+
+Cap each search round (e.g., 15 min). When obvious queries dry up, ask **what is conspicuously absent**: a topic with no critical papers since 2022 may indicate (a) consensus, (b) abandonment, or (c) missing index — investigate which.
+
+---
+
+## 6. Source Stratification by Domain
+
+(Authority hierarchy + traps. Full per-source tables in `source-directory.md`.)
+
+### 6.1 Software / Web Engineering
+
+**Hierarchy**: standards bodies (W3C, WHATWG, ECMA, IETF RFC) > vendor official docs (MDN, Microsoft Learn, AWS/GCP/Azure docs) > package registries with provenance (npm, PyPI, crates.io) > GitHub READMEs of the source repo > Stack Overflow (vote-weighted) > blog posts > AI-generated tutorials.
+
+- **MDN Web Docs**. <https://developer.mozilla.org/>. Mozilla-maintained, browser-feature-status authoritative.
+- **WHATWG Living Standards** (HTML, DOM, Fetch, URL). <https://spec.whatwg.org/>.
+- **ECMAScript** (TC39). <https://tc39.es/ecma262/>. JS language spec.
+- **RFC Editor** (IETF). <https://www.rfc-editor.org/>. Network protocols.
+
+**Traps**: cargo-culted `npm install` advice from outdated tutorials; AI-rewritten Stack Overflow answers; deprecation cascades not reflected in popular blogs.
+
+### 6.2 UX/Design
+
+**Hierarchy**: Baymard (commerce-UX gold standard, paywalled) > NN/g > IxDF > design-system canonical sources (Material 3, Apple HIG, Microsoft Fluent, Shopify Polaris, IBM Carbon, Atlassian Design System) > W3C WAI / ARIA Authoring Practices > Smashing/A List Apart > Medium/dev.to.
+
+- **Baymard Institute**. <https://baymard.com/research>.
+- **Nielsen Norman Group**. <https://www.nngroup.com/>.
+- **Interaction Design Foundation** (IxDF). <https://www.interaction-design.org/literature>.
+- **W3C WAI ARIA Authoring Practices**. <https://www.w3.org/WAI/ARIA/apg/>.
+- **Deque University**. <https://dequeuniversity.com/>. Accessibility specialists.
+
+### 6.3 Academic
+
+**Hierarchy**: peer-reviewed journals (verified via Crossref DOI) > arXiv/bioRxiv preprints (note "preprint", check if peer-reviewed since) > conference proceedings (top-tier: ACM/IEEE/Springer LNCS) > theses (institutional) > working papers > ResearchGate-only PDFs.
+
+- **Crossref API**. <https://api.crossref.org/>. Resolve DOI → metadata. `https://api.crossref.org/works/{doi}` returns JSON.
+- **DOI**. <https://www.doi.org/>. Persistent identifier; resolve via `https://doi.org/{doi}`.
+- **ORCID**. <https://orcid.org/>. Author identifier; verify name + affiliation.
+- **Google Scholar**, **Semantic Scholar**, **Scopus**, **Web of Science**.
+- **arXiv** <https://arxiv.org/>, **PubMed** <https://pubmed.ncbi.nlm.nih.gov/>.
+
+### 6.4 Business / Market
+
+**Hierarchy**: tier-1 analyst firms (Gartner, Forrester, IDC) > management consultancies (McKinsey, BCG, Bain) > business journals (HBR, MIT Sloan Mgmt Review) > trade press (CB Insights, a16z research) > startup data (PitchBook, Crunchbase) > general business news.
+
+### 6.5 News / Current Events
+
+**Hierarchy**: wire services (Reuters, AP, AFP) > newspapers of record (NYT, WSJ, FT, The Economist) > regional papers > digital natives (Axios, The Verge for tech) > aggregators > social media.
+
+- **Wayback Machine**. <https://web.archive.org/>. Verify what a URL said on a given date.
+- **Snopes** <https://www.snopes.com/>, **PolitiFact** <https://www.politifact.com/>, **FactCheck.org** <https://www.factcheck.org/>.
+
+### 6.6 Standards Bodies
+
+**ISO** <https://www.iso.org/> (paywalled, cite by ISO/IEC NNNN-N:YYYY), **IEEE** <https://standards.ieee.org/>, **NIST** <https://www.nist.gov/publications> (free), **W3C** <https://www.w3.org/TR/>, **IETF** <https://datatracker.ietf.org/>, **OASIS** <https://www.oasis-open.org/>, **ECMA** <https://ecma-international.org/publications-and-standards/>.
+
+Per-domain authenticity protocols are in `source-directory.md`.
+
+---
+
+## 7. Freshness Calibration by Content Type
+
+### 7.1 Half-Lives
+
+| Content type                                                                               | Useful life  | Recheck cadence  |
+| ------------------------------------------------------------------------------------------ | ------------ | ---------------- |
+| **Fast-moving** (frontend frameworks, AI/LLM SOTA, cloud pricing, browser support)         | Weeks–months | Every 1–3 months |
+| **Medium** (established libraries, design patterns, UX best practices, security baselines) | 1–2 years    | Annually         |
+| **Slow** (language fundamentals, CS theory, HCI canonical research)                        | 5+ years     | Every 5y         |
+| **Near-permanent** (math, physics, history)                                                | Decades      | When citing      |
+
+### 7.2 "Fresh-Looking but Stale"
+
+- **Republished date**: blog updated `2025-04-01` but content untouched since 2021 — check archived versions on Wayback.
+- **Link rot**: cited URL 404s; the cited claim cannot be verified — mark `[link-rot]`.
+- **Phantom updates**: CMS bumps the modified-at on every save with no content change.
+
+Detection: diff Wayback snapshots; check if cited libraries/versions are current; sniff for "2025" timestamps with 2020-era code samples.
+
+### 7.3 "Stale but Evergreen"
+
+- Foundational HCI: Fitts (1954), Hick-Hyman (1952, 1953), Miller's 7±2 (1956) — still apply.
+- CS classics: Knuth, Dijkstra, Lamport — cite freely.
+- Standards that have not been updated _because they're stable_: RFC 1918 (private IP space, 1996) — current.
+
+Heuristic: a primary source from 1990 cited by 2024 textbooks in the same field is evergreen.
+
+### 7.4 Breaking-Change Detection
+
+- **Major version bumps**: read CHANGELOG / release notes.
+- **Deprecation notices**: search `site:<vendor-docs> deprecated <feature>`.
+- **CVE feeds**: <https://nvd.nist.gov/> and <https://github.com/advisories> for security-relevant freshness.
+- **Browser feature status**: <https://caniuse.com/>, <https://chromestatus.com/features>, <https://developer.mozilla.org/en-US/docs/Web/API> "Browser compatibility" tables.
+
+---
+
+## 8. Anti-Hallucination Protocols
+
+### 8.1 Why This Matters
+
+LLMs fabricate plausibly. Evidence:
+
+- **Bender EM, Gebru T, McMillan-Major A, Shmitchell S.** "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" _FAccT '21_ 2021. doi:10.1145/3442188.3445922.
+- **Walters WH, Wilder EI.** "Fabrication and errors in the bibliographic citations generated by ChatGPT." _Sci Rep_ 2023;13:14045. doi:10.1038/s41598-023-41032-5. Found that 47% of GPT-3.5 citations were fabricated.
+- **Alkaissi H, McFarlane SI.** "Artificial Hallucinations in ChatGPT: Implications in Scientific Writing." _Cureus_ 2023;15(2):e35179. doi:10.7759/cureus.35179.
+
+### 8.2 Verification Pipeline (the URL+QUOTE+ACCESSED-AT+VERIFY-METHOD protocol)
+
+For every claim:
+
+1. **URL resolves**: HTTP HEAD or GET returns 200 (or 301 → 200). Capture final URL.
+2. **DOI exists in Crossref**: `GET https://api.crossref.org/works/{doi}` returns 200 and metadata fields match the cited title/author/year.
+3. **Quote-in-source grep**: extract a 6–12-word verbatim quote from the source page; the synthesis must include that quote and the page must contain it. The `verify-claims.sh` script grep-confirms every quote.
+4. **Author verification**: ORCID lookup or institutional page link; for non-academic, byline + masthead match.
+5. **Date verification**: distinguish "publication date" from "last-updated"; both should be plausible. If only updated date is shown, search for the original publication date in archives.
+6. **Cross-citation chain**: walk back to the primary source. If an article cites a paper, the _paper_ must verify, not just the citing article.
+7. **Screenshot evidence**: for time-sensitive or paywalled content, save a Playwright screenshot with timestamp.
+
+### 8.3 Sources That Only Exist in LLM Outputs
+
+A common failure mode: LLM cites a paper that does not exist. Detection:
+
+- Crossref returns 404
+- Google Scholar returns no result for `"<exact title>"`
+- Author has no other publications matching
+- Journal name slightly garbled ("Journal of Data Science" vs "Journal of the American Statistical Association")
+- DOI structure invalid (Crossref DOIs follow `10.\d{4,9}/.+`)
+
+### 8.4 Citogenesis Defense
+
+The citogenesis loop (§4.2) is the deepest hallucination trap. Defense:
+
+- **Find the earliest publication date** for any claim — Wayback Machine, archived versions.
+- If the earliest source is itself uncited, the claim has no basis and should be marked `[unverified]`.
+
+### 8.5 Spot-Check Protocol
+
+After synthesis, the research-verify agent randomly samples 10% of claims and runs the full pipeline. Any failure escalates: rerun query, mark contested, or remove.
+
+---
+
+## 9. Knowledge Base Management
+
+### 9.1 Deduplication
+
+- **Content hashing**: SHA-256 of normalized text (lowercased, whitespace-collapsed, stop-words removed) — exact duplicate detector.
+- **Semantic similarity**: cosine over sentence embeddings (e.g., MiniLM-L6, ~384 dim); threshold 0.85+ flags near-duplicates.
+- **Citation overlap**: documents sharing > 60% of citations are likely covering the same ground.
+
+### 9.2 Versioning
+
+Each `/docs/research/<topic>.md` carries frontmatter:
+
+```yaml
+created: 2025-11-01
+updated: 2026-04-25
+freshness: fresh # fresh | aging | stale | outdated
+last-verified: 2026-04-25
+sources-checked: 18
+```
+
+Update `updated` on every change; `last-verified` only when claims are re-grep-checked.
+
+### 9.3 Cross-Linking
+
+- `[[concept-slug]]` for outbound wikilinks.
+- `index.md` per folder + `/docs/index.md` root, both with backlink registries.
+- Tag taxonomy in frontmatter (`tags: [react, data-fetching, ssr]`) — controlled vocabulary maintained in `/docs/research/_tags.md`.
+
+### 9.4 Stale Detection
+
+Cron-style review: any doc older than its content-type half-life (§7) is flagged. Stale ≠ delete — re-verify or annotate.
+
+### 9.5 Research Debt
+
+Unverified claims, broken links, and contested claims accumulate. Track in `/docs/research/_debt.md`. Treat like tech debt: pay down, do not ignore.
+
+### 9.6 ADRs
+
+**Nygard MT.** "Documenting Architecture Decisions." 2011. <https://cognitect.com/blog/2011/11/15/documenting-architecture-decisions>. Adopted across the industry. Format: Title, Status, Context, Decision, Consequences. The skill emits ADRs to `/docs/decisions/NNNN-decision.md`.
+
+### 9.7 Note-Taking Frameworks
+
+- **Zettelkasten**. Niklas Luhmann's slip-box system. Atomic notes, dense linking, no folders. **Ahrens S.** _How to Take Smart Notes_. CreateSpace, 2017. ISBN 978-1542866507.
+- **PARA**. **Forte T.** _Building a Second Brain_. Atria Books, 2022. ISBN 978-1982167387. Projects / Areas / Resources / Archive.
+- **Evergreen notes**. **Matuschak A.** <https://notes.andymatuschak.org/Evergreen_notes>. Notes that are written for the future self, refactored over time.
+- **MOC** (Map of Content). **Milo N.** _Linking Your Thinking_. <https://www.linkingyourthinking.com/>. Index notes that organize related atomic notes.
+
+The skill borrows: atomic note granularity (Zettelkasten), evergreen rewriting on update (Matuschak), MOC pattern (`/docs/research/index.md`).
+
+---
+
+## 10. Output Formats
+
+### 10.1 Human Markdown (Default)
+
+GitHub-flavored. Frontmatter + sections + inline citations `[Author Year]`. The skill's `/docs/research/<topic>.md` template uses this.
+
+### 10.2 JSON-LD
+
+```json
+{
+	"@context": "https://schema.org",
+	"@type": "ScholarlyArticle",
+	"name": "topic title",
+	"datePublished": "2026-04-25",
+	"citation": [{ "@id": "https://doi.org/10.1136/bmj.n71" }]
+}
+```
+
+JSON-LD 1.1: <https://www.w3.org/TR/json-ld11/>.
+
+### 10.3 ADR
+
+See §9.6.
+
+### 10.4 RDF / Turtle
+
+```turtle
+@prefix skos: <http://www.w3.org/2004/02/skos/core#> .
+@prefix : <https://docs.example.com/concepts/> .
+
+:react-server-components a skos:Concept ;
+  skos:prefLabel "React Server Components"@en ;
+  skos:broader :react-rendering-models ;
+  skos:related :streaming-ssr .
+```
+
+### 10.5 OWL
+
+Use only when reasoning is needed (rare in this skill). Protégé export.
+
+### 10.6 SKOS Concept Schemes
+
+Encoded as `concepts:` block in frontmatter — see `ontology-patterns.md`.
+
+### 10.7 BibTeX / CSL-JSON
+
+For academic citations, the skill outputs both:
+
+```bibtex
+@article{page2021prisma,
+  doi = {10.1136/bmj.n71},
+  author = {Page, Matthew J and others},
+  journal = {BMJ},
+  year = {2021},
+  volume = {372},
+  pages = {n71}
+}
+```
+
+CSL-JSON: <https://docs.citationstyles.org/en/stable/specification.html>.
+
+### 10.8 Citation Styles
+
+- **APA 7th ed.**: <https://apastyle.apa.org/>.
+- **MLA 9th ed.**: <https://style.mla.org/>.
+- **Chicago 17th ed.**: <https://www.chicagomanualofstyle.org/>.
+
+### 10.9 Mermaid
+
+For ontology slices and decision trees in markdown — works in GitHub, Obsidian, Notion. <https://mermaid.js.org/>.
+
+```mermaid
+graph LR
+  A[Topic] --> B[Concept 1]
+  A --> C[Concept 2]
+  B -- depends-on --> C
+```
+
+### 10.10 Obsidian Frontmatter and Progressive Disclosure
+
+```yaml
+---
+title: React Data Fetching
+aliases: [data fetching in react, RDF-react]
+tags: [react, data-fetching]
+concepts: [server-components, suspense, streaming]
+freshness: fresh
+---
+```
+
+Progressive disclosure: TL;DR → Key claims → Detail → Sources → Appendix. Readers (human and LLM) get value at every level.
+
+---
+
+## 11. Orchestration
+
+### 11.1 Four-Agent Architecture
+
+| Agent                   | Model  | Role                                                                                            |
+| ----------------------- | ------ | ----------------------------------------------------------------------------------------------- |
+| **research-scout**      | Haiku  | Decompose question into sub-queries, check `/docs/research/` cache, build ontology map skeleton |
+| **research-query**      | Sonnet | Run parallel WebSearch / context7 / fetch searches, extract claims with URL+QUOTE+ACCESSED-AT   |
+| **research-synthesize** | Sonnet | Triangulate, dedupe, build ontology, write `/docs/research/<topic>.md`                          |
+| **research-verify**     | Haiku  | Anti-hallucination grep (URL resolves, DOI in Crossref, quote-in-source)                        |
+
+### 11.2 Parallelization Rules
+
+- Independent sub-queries → parallel `research-query` invocations (cap concurrency at ~5).
+- Synthesis is sequential (depends on all queries).
+- Verification can parallelize over claims (one HEAD/GET per URL).
+
+### 11.3 Token Budget
+
+- Scout: ≤ 5k tokens (decomposition only).
+- Query (per sub-query): ≤ 30k tokens (search + extract).
+- Synthesize: ≤ 80k tokens (sees all extracted claims, writes output).
+- Verify: ≤ 20k tokens (sees claims + URLs, runs checks).
+
+If a topic exceeds total budget, the scout splits into multiple research sessions (`/docs/research/topic-part-1.md`, `topic-part-2.md`).
+
+---
+
+## 12. Domain Playbooks
+
+Pointer only — full playbooks live in `domain-playbooks.md`. Coverage:
+
+1. UX/design pattern research (Baymard-style competitive analysis, NN/g 10 heuristics, accessibility audit plan)
+2. Library / framework evaluation (community health, bundle size, ecosystem fit, migration cost)
+3. API integration research (auth, rate limits, SDK quality, pricing tiers, SLA)
+4. Architectural decision research (ADR scaffolding, RFC prior art, trade-off matrix, reversibility)
+5. Market / competitive research (TAM/SAM/SOM, positioning quadrants, Porter 5F)
+6. Academic literature review (PRISMA-ScR scoping flow, inclusion/exclusion templates)
+7. News & current-events research (wire-source-first, timeline construction, AllSides/Ad Fontes bias)
+8. Security research (CVE/NVD lookup, advisory chains, vendor disclosure timing)
+9. Pricing & cost research (vendor tiers, hidden fees, TCO, deprecation pricing risk)
+
+---
+
+## 13. Confidence and Epistemic Humility
+
+### 13.1 Confidence Levels
+
+| Level         | Criteria                                                                                                                    |
+| ------------- | --------------------------------------------------------------------------------------------------------------------------- |
+| **High**      | ≥ 3 independent primary sources, all agree, fresh per content-type half-life, authoritative authorities (tier-1 in §6)      |
+| **Medium**    | 2 primary + 1+ corroborating; or 3 primary with minor disagreements; or 1 authoritative primary that is the standard itself |
+| **Low**       | 1 primary with no corroboration; or sources disagree; or freshness exceeded; or only secondary sources                      |
+| **Contested** | ≥ 2 primaries that disagree — record both positions with attribution, no synthesis claim                                    |
+
+### 13.2 Known Unknowns
+
+Explicitly list what was searched for and not found. This is itself evidence (e.g., "no published benchmarks for X under Y conditions as of 2026-04").
+
+### 13.3 Unknown Unknowns
+
+Adapted from Rumsfeld DH. Press briefing, US DoD, 2002-02-12. <https://archive.ph/20180320091111/http://archive.defense.gov/Transcripts/Transcript.aspx?TranscriptID=2636>. The research output should briefly note adjacent areas the researcher _did not_ explore and why — surfaces blind spots for the reader.
+
+### 13.4 Disagreement Tracking
+
+`/docs/research/<topic>.md#disagreements` section. Each disagreement: positions, sources per position, possible resolutions, current verdict ("contested").
+
+### 13.5 Dead Ends
+
+Document searches that returned nothing useful — saves a future re-run. `/docs/research/_dead-ends.md`.
+
+### 13.6 Replication Crisis
+
+- **Ioannidis JPA.** "Why Most Published Research Findings Are False." _PLoS Med_ 2005;2(8):e124. doi:10.1371/journal.pmed.0020124.
+- **Open Science Collaboration.** "Estimating the reproducibility of psychological science." _Science_ 2015;349(6251):aac4716. doi:10.1126/science.aac4716. Replicated 36 of 100 psychology findings.
+- **Camerer CF, Dreber A, Forsell E, et al.** "Evaluating replicability of laboratory experiments in economics." _Science_ 2016;351(6280):1433–1436. doi:10.1126/science.aaf0918.
+
+Implication: a single empirical paper, however peer-reviewed, is not high-confidence on its own. Prefer pre-registered, replicated, or meta-analytic evidence.
+
+### 13.7 Bayesian Update
+
+Treat new evidence as a posterior shift, not a binary truth flip. Format the update explicitly:
+
+> Prior: confidence Medium based on Smith 2022.
+> New: Jones 2025 contradicts on subgroup X.
+> Posterior: confidence Medium-Low; flagged contested for subgroup X, Medium elsewhere.
+
+---
+
+## 14. Ethics and Bias
+
+### 14.1 Source Bias
+
+| Bias type    | Source                              | Mitigation                                  |
+| ------------ | ----------------------------------- | ------------------------------------------- |
+| Commercial   | Vendor-funded "research"            | Read disclosures; weigh against independent |
+| Political    | Partisan think-tanks                | Use AllSides / Ad Fontes Media bias rating  |
+| Academic     | Field orthodoxy / publication norms | Look for cross-field corroboration          |
+| Geographical | English-only or West-centric        | Search non-English primaries when relevant  |
+| Funding      | Industry-sponsored studies          | Always check Acknowledgments                |
+
+### 14.2 Sampling Bias
+
+Convenience samples (especially online surveys) over-represent always-online populations. Cite sample frame explicitly.
+
+### 14.3 Publication Bias / File-Drawer Effect
+
+- **Rosenthal R.** "The 'file drawer problem' and tolerance for null results." _Psychol Bull_ 1979;86(3):638–641. doi:10.1037/0033-2909.86.3.638. Negative results disproportionately go unpublished.
+- Mitigation: search trial registries (ClinicalTrials.gov, OSF preregistrations) for studies that started but never published.
+
+### 14.4 Survivorship Bias
+
+Studies of "successful X" without a control group of failed X cannot identify success factors. Classic case: Wald A's WWII bullet-hole analysis. Mangel M, Samaniego FJ. "Abraham Wald's work on aircraft survivability." _J Am Stat Assoc_ 1984;79(386):259–267. doi:10.1080/01621459.1984.10478038.
+
+### 14.5 Confirmation Bias
+
+- **Popper K.** _The Logic of Scientific Discovery_. 1959. Falsification > confirmation.
+- Practice: for every hypothesis, search for evidence _against_ it before evidence _for_. Bake into the query plan.
+
+### 14.6 AI Content Laundering
+
+Generated content republished as human-authored, then cited. Detect: §4.7 red flags + §8.3 checks.
+
+### 14.7 Citation Washing
+
+Citing a high-status source for a claim that source did not actually make. Detect: quote-in-source grep (§8.2).
+
+### 14.8 Proprietary vs Open
+
+Paywalled research (Gartner, Forrester, ISO) cannot be redistributed. Quote sparingly under fair use, link, and note paywall. Open-access (PLOS, eLife, NIST, RFCs, W3C) can be quoted at length.
+
+---
+
+## 15. Concrete Skill Architecture
+
+The `research` skill is implemented as 4 specialist agents (research-scout, research-query, research-synthesize, research-verify) coordinated by a top-level SKILL.md, supported by 5 reusable scripts (search dispatch, claim extraction, ontology builder, claim verifier, freshness checker), 5 markdown templates (topic root, finding-level, ADR, scoping-review summary, dead-end log), and 4 reference docs (`research-methodology.md` — this file; `ontology-patterns.md`; `source-directory.md`; `domain-playbooks.md`). Output lands in `/docs/research/<topic>.md` with sibling `index.md` and `_tags.md` indexes.
+
+The pipeline mirrors the sibling `super-design` and `e2e-audit` skills: source-first discovery, evidence-anchored output (URL+QUOTE+ACCESSED-AT+VERIFY-METHOD), report-then-act gating, idempotent re-runs that only re-verify what changed since last run. Every claim in every output file must round-trip through `verify-claims.sh` before commit.