Wearly — Evidence and References¶

The single source of truth for what informs Wearly's recommendation logic. Every rule pack and reasoning line in the agent traces back to one of the source categories named below. Where a specific citation has been verified, it appears in that section. Where a category is informed by general practice or by frameworks that haven't been individually verified yet, that's stated plainly with a [to verify] marker.

This document never invents citations.

1. The honesty contract¶

Wearly's documentation, the agent's reasoning trail, and this references document all hold themselves to four rules:

Styling is not an exact science. Recommendations are evidence-informed, design-informed, and user-preference-driven — never described as "correct," "objectively right," or "optimal."
Body-positive language is non-negotiable. Implemented as a runtime contract in fit_tool.check_reasoning_for_forbidden_language() (CI-checked) and now as a static documentation check (tests/test_documentation_integrity.py). Forbidden tokens: flaw / fix / hide / minimize / correct / problem area / slimming / shameful.
No fabricated citations. Categories may be named when they're textbook common knowledge. Specific papers, authors, and findings appear only after verification.
The user is the final reviewer. The reject-and-regenerate loop exists precisely because no rule system can be authoritative about personal taste.

2. How to read this document¶

Each source category below carries six items:

Field	What it means
Definition	One-sentence scope of the category.
Why it matters for Wearly	Which rule pack(s) draw on this category.
Source types to gather	Specific kinds of sources a researcher should seek (search queries, venues, frameworks).
Current basis	What Wearly's rules are informed by today — general practice, well-known framework, or verified citation.
Verification status	One of: `general practice`, `well-known framework`, `verified citations`, `to verify`. After the May 2026 verified-sources pass, all nine categories now carry `verified citations`.
Pending items	A short list of specific things still needing verification.

This pattern lets a future contributor pick any rule pack, find the relevant categories, and either rely on what's there or pick up the verification work.

3. Source categories¶

3.1 Fashion recommender systems¶

Definition. The subfield of recommender systems that addresses garment selection, outfit completion, and personalized fashion suggestions.

Why it matters for Wearly. Underlies the seven-step workflow as a whole, especially Step 4 (candidate filtering), Step 5 (outfit construction), and the reject-and-regenerate loop.

Source types to gather. - Survey papers on recommender systems (collaborative filtering vs. content-based vs. hybrid). - Conference proceedings from ACM RecSys, SIGIR, KDD, IEEE-related venues. - Fashion-specific recommendation work — search terms: "outfit compatibility model," "fashion outfit recommendation," "set-based recommendation," "garment retrieval." - The "Recommender Systems Handbook" (Ricci, Rokach, Shapira) is the standard reference text. [to verify edition]

Current basis. Rule-based / content-based filtering. Wearly does not use collaborative filtering today (single-user scope today, no cross-user signal). The rule layer is intentional — see book/04-styling-knowledge-base.md for why the reasoning core is inspectable rules.

Verified citations.

Cited: Ricci, F., Rokach, L., & Shapira, B. (Eds.). (2022). Recommender Systems Handbook (3^rd ed.). Springer. DOI: 10.1007/978-1-0716-2197-4. Claim it supports: Recommender systems are a mature field with content-based, collaborative-filtering, and hybrid approaches; Wearly's content-based rule layer is one of the named families. Relevant rule(s): the seven-step workflow as a whole; wardrobe-filtering-rules.md.

Verification status. verified citations for the recommender-systems framing; [to verify] for fashion-specific explainable-recommender user studies.

Pending items. - A user-study paper on explanation interfaces for fashion or e-commerce recommenders.

3.2 Outfit compatibility¶

Definition. Methods for assessing whether two or more garments form a visually coherent outfit (style, color, formality, season, occasion).

Why it matters for Wearly. Drives Step 5's dress-vs-separates branching, the formality-matching heuristics, and the rule that a gym outfit should never receive a formal coat.

Source types to gather. - Compatibility-prediction papers (often graph-based or pairwise-similarity-based). - Polyvore-dataset publications — a widely-cited public dataset of curated outfits. [to verify availability] - Design pedagogy texts on outfit construction.

Current basis. Industry styling heuristics encoded as rules in styling_agent.py (REQUIRED_PIECES, DRESS_OCCASIONS, formality thresholds). No learned compatibility model. The verified citations below name the alternative learned approach — they're cited to position Wearly against it, not to claim Wearly implements it.

Verified citations.

Cited: Han, X., Wu, Z., Jiang, Y.-G., & Davis, L. S. (2017). Learning Fashion Compatibility with Bidirectional LSTMs. In Proceedings of the 25^th ACM International Conference on Multimedia (MM '17). arXiv:1707.05691. Source for the widely-cited Polyvore outfit dataset. Claim it supports: Outfit compatibility has been formalized as a learnable sequence task on real curated-outfit data. Wearly's rule layer is one alternative; this paper names the learned-model alternative we deliberately did not pursue. Relevant rule(s): occasion-rules.md R3 (dress-vs-separates branching), book/04-styling-knowledge-base.md ("Why the reasoning core is inspectable rules").

Cited: Vasileva, M. I., Plummer, B. A., Dusad, K., Rajpal, S., Kumar, R., & Forsyth, D. (2018). Learning Type-Aware Embeddings for Fashion Compatibility. In Proceedings of the European Conference on Computer Vision (ECCV 2018). DOI: 10.1007/978-3-030-01270-0_24. arXiv:1803.09196. Claim it supports: Item-type-aware embeddings improve outfit-compatibility prediction. Reinforces that type (top / bottom / shoes / accessory) is a meaningful axis — the same axis Wearly's REQUIRED_PIECES map uses, only as an inspectable lookup. Relevant rule(s): occasion-rules.md R2 (required piece-types).

Verification status. verified citations.

Pending items. - A peer-reviewed paper or industry whitepaper describing a deployed outfit-compatibility model in a real product (rather than a benchmark study).

3.3 Color harmony¶

Definition. Principles for combining colors so they read as coherent and intentional. Spans both perceptual science (color contrast, simultaneous contrast) and design conventions (analogous, complementary, triadic schemes, seasonal-color analysis).

Why it matters for Wearly. Step 7 scores every outfit against color_rules.json. The named-color palettes in color_tool.py are themselves color-harmony decisions.

Source types to gather. - Foundational color-theory texts. Well-known frameworks worth verifying: - Johannes Itten's color wheel and contrast principles. A foundational reference in design education; the specific work to cite is "The Art of Color" / "The Elements of Color." [to verify exact edition + chapter] - Munsell color system. Standard for systematic color specification. [to verify] - CIE color spaces (CIELAB, CIELUV). International standards for color measurement. [to verify standard numbers] - Seasonal-color analysis literature. Note: the "Color Me Beautiful" framework (Carole Jackson) is industry/commercial; it's widely referenced but is not peer-reviewed science. Should be cited as industry practice, not science. - Perceptual psychology papers on color preference and skin-tone interaction. [to verify]

Current basis. Three palettes (warm olive, cool fair, deep warm) informed by general warm/cool color-theory principles and industry practice. The +8 / +4 / −10 scoring formula is design-informed, not derived from a specific empirical model.

Verified citations. Each names a foundational framework Wearly's palettes draw on as background. Wearly does NOT compute CIE distances or Munsell coordinates at runtime — these are the systems the warm/cool convention itself descends from.

Cited: Itten, J. (1961). Kunst der Farbe. Otto Maier Verlag, Ravensburg. English condensation: Itten, J., & Birren, F. (Ed.). (1970). The Elements of Color. Van Nostrand Reinhold / John Wiley & Sons. ISBN 0-442-24038-4. Claim it supports: A formal vocabulary of color contrasts (hue, light/dark, cold/warm, complementary, simultaneous, saturation, extension) underlies design-school color teaching. The warm/cool axis Wearly's three palettes pivot on comes from this tradition. Relevant rule(s): color-coordination-rules.md R1 (skin-tone palettes).

Cited: Munsell, A. H. (1905). A Color Notation. Geo. H. Ellis Co., Boston. Public domain text available via Project Gutenberg, eBook #26054. Claim it supports: A measured color system (hue, value, chroma) predates Wearly's named colors by a century and is the basis of subsequent standardization. Wearly's named palette is intentionally a human-readable layer above a system like this — not a replacement for one. Relevant rule(s): color-coordination-rules.md R1.

Cited: ISO 11664-4:2008 (CIE S 014-4:2007). Colorimetry — Part 4: CIE 1976 Lab Colour space. International Organization for Standardization. iso.org/standard/52497.html. *Claim it supports:** A perceptually-oriented color space (CIELAB) exists as an international standard. Wearly's prototype does not compute ΔE distances; this citation is acknowledgement of the framework, not a claim that Wearly uses it. Relevant rule(s): color-coordination-rules.md (Source basis only — not invoked at runtime).

Verification status. verified citations for the foundational color frameworks; [to verify] for any peer-reviewed empirical study of perceived flatteringness across skin-tone × garment-color combinations.

Pending items. - A peer-reviewed empirical study, if any exists, of perceived flatteringness across skin-tone × garment-color combinations. Note: the "Color Me Beautiful" framework is industry/commercial and is explicitly NOT cited as science.

3.4 Body-shape-aware / fit-aware styling¶

Definition. Guidance on garment cut, drape, and proportion for varying body proportions, fit preferences, and modesty needs.

Why it matters for Wearly. Informs skills/wearly-styling-agent/fit-silhouette-rules.md and fit_tool.fit_alignment_notes().

Source types to gather. - Garment-construction and pattern-making texts (apparel-design textbooks). - Wearable-fit literature from human-factors / ergonomics venues. - Inclusive-fashion research — search terms: "adaptive clothing," "inclusive sizing," "size-inclusive design."

Current basis. Industry styling heuristics + user-declared preferences. This is the area where Wearly is most deliberately careful — see §4 below for the framing decision.

Verified citations.

Cited: Hokka, J. (2024). Gender and the Diversity of the Human Body as Challenges for the Inclusive Design of Wearable Technology. Fashion Practice, 16(1) (published online September 2023). DOI: 10.1080/17569370.2023.2250153. Claim it supports: Body diversity is a documented challenge for inclusive design; treating a small set of shape categories as universal is a known limitation. Supports Wearly's framing of body_shape as a user-declared proportion preference, not a classification we perform. Relevant rule(s): fit-silhouette-rules.md R2 (optional fields), R7 (body-shape framing); §4 of this document.

Verification status. verified citations for the inclusive-design framing; general practice for the proportion-based styling heuristics themselves. The body-shape category labels (hourglass / pear / apple / inverted triangle / rectangle) used in wardrobe.json's owner.body_shape field remain industry heuristics, not a scientific taxonomy. They're useful as proxies for proportion-related styling rules; they're not a claim about human biology.

Pending items. - A peer-reviewed critique of common body-shape categorization schemes (the field has criticism in academic literature but a direct verified citation has not yet been added).

3.5 Explainable recommendation systems (XAI for recommenders)¶

Definition. Techniques and frameworks for making recommendation systems' decisions inspectable, justifiable, and accountable to the user.

Why it matters for Wearly. The entire result["reasoning"] array exists to satisfy this concern. Every Step 5 pick gets a sentence. Every freshness / fit-alignment note appears with its rule label. The reject-and-regenerate loop is XAI in action — the user argues with the system and the system responds.

Source types to gather. - XAI surveys, particularly for recommender systems. - Work on counterfactual and example-based explanations for recommendation. - Survey: Tim Miller's work on explanation in AI is widely cited (early 2019 publication on the social-science-grounded view of explanations). [to verify exact citation] - HCI / IUI venue papers on user studies of explanation interfaces.

Current basis. Wearly's explanation style is natural-language template — one or more sentences per decision, with optional rule citations. Aligns with the established XAI position that explanations should be (a) faithful, (b) selective, and © socially appropriate.

Verified citations.

Cited: Miller, T. (2019). Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267, 1–38. DOI: 10.1016/j.artint.2018.07.007. Preprint: arXiv:1706.07269. Claim it supports: Explanations are contrastive, selective, and socially appropriate rather than full causal traces. Directly supports Wearly's design choice to surface one-line reasons per decision (not a complete derivation) and to make rejection-driven regeneration a first-class loop. Relevant rule(s): the result["reasoning"] array as a whole; the rule-citation chain (§6a); wardrobe-filtering-rules.md R5 (rejection exclusion).

Verification status. verified citations.

Pending items. - A user-study paper specifically on explanation interfaces for fashion or e-commerce recommenders.

3.6 Human-centered AI / personalization¶

Definition. Design principles for AI systems that augment rather than replace human judgment, respect user autonomy, and surface meaningful agency.

Why it matters for Wearly. Informs the whole product stance — the reject-and-regenerate loop, the "Wear this outfit today" confirmation, the editable fit profile, the manual override available at every form field, the body-positive language contract.

Source types to gather. - The Microsoft AI Guidelines for Human-AI Interaction are publicly available and widely referenced. [to verify URL + version] - The Google PAIR (People + AI Research) Guidebook is publicly available. [to verify URL + version] - HAI / Stanford HAI publications. - Research on mixed-initiative systems and AI agency / user control trade-offs.

Current basis. Wearly was explicitly designed around the principle "an agent" — phrasing taken from the SEIS 666 course materials and consistent with the human-centered-AI school of thought. Every recommendation is a starting point the user can push back on.

Verified citations.

Cited: Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for Human-AI Interaction. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (paper 3, pp. 1–13). DOI: 10.1145/3290605.3300233. Project page: microsoft.com/en-us/research/project/guidelines-for-human-ai-interaction. Claim it supports: 18 design guidelines for AI-infused products covering "initially," "during interaction," "when wrong," and "over time." Wearly's reject-and-regenerate loop, the editable fit profile, the body-positive language contract, and the never-auto-submit confirmations are direct expressions of guidelines such as "support efficient correction," "make clear what the system can do," and "remember recent interactions." Relevant rule(s): the agentic stance; wardrobe-filtering-rules.md R5; privacy-guidelines.md R5 (user pushback is authoritative).

Cited: Google PAIR (People + AI Research). People + AI Guidebook. pair.withgoogle.com/guidebook. Open-access design guidance for human-centered AI products. Claim it supports: Practical patterns for user-needs framing, mental-model alignment, explainability, feedback, errors, and trust calibration in AI products. Wearly's UX (the seven-step reasoning trail, the wardrobe-gap "honest gaps over forced fits" stance, the calendar/weather context surfaces) aligns with the Guidebook's mental-model and trust-calibration chapters. Relevant rule(s): shopping-gap-rules.md R5; the reasoning trail as a whole.

Verification status. verified citations.

Pending items. (none for this category at present)

3.7 Wardrobe management and digital closet systems¶

Definition. Tools and research that help users catalog, manage, and reason about their owned garments — distinct from shopping recommenders.

Why it matters for Wearly. The wardrobe model (wardrobe.json + user_wardrobe.json overlay), the Wardrobe Builder, the wear-history rotation logic, and the wardrobe-gap detection all draw on this category.

Source types to gather. - Academic work on sustainable fashion / wardrobe utilization — sometimes appears in CHI venues, sometimes in design-research journals. - Industry whitepapers or case studies from established closet-management apps (Acloset, Stylebook, OpenWardrobe, Smart Closet). These are competitive references, not academic, and should be cited as industry practice. - Wear-history rotation has both academic precedent (resource-allocation, scheduling) and consumer-fashion-blog precedent.

Current basis. Wear-history rotation formula (frequency penalty + recency penalty, 0.4 floor) is design-informed — see skills/wearly-styling-agent/wear-history-rules.md. The 0.4 floor is a deliberate values choice: a much-loved piece is never exiled.

Verification status. general practice.

Pending items. - Academic literature on consumer wardrobe utilization, if any exists. Likely terms: "wardrobe-30," "Project 333," "cost-per-wear analysis." - Industry case studies on closet-management product behaviors.

3.8 Personalization and user feedback¶

Definition. Techniques for incorporating explicit and implicit user feedback into recommendation quality.

Why it matters for Wearly. The reject-and-regenerate loop is explicit negative feedback. The "Wear this outfit today" button is explicit positive feedback. The fit profile is declarative preference. All three converge in Step 5's pool sorting.

Source types to gather. - Recommender-systems literature on explicit vs. implicit feedback. - HCI literature on preference elicitation. - Active-learning / cold-start research relevant to a fresh user's first outfit.

Current basis. Wearly uses explicit feedback only — no implicit-click-tracking, no behavioral inference. This is a deliberate privacy choice (see §3.9).

Verified citations.

Cited: Hu, Y., Koren, Y., & Volinsky, C. (2008). Collaborative Filtering for Implicit Feedback Datasets. In Proceedings of the 8^th IEEE International Conference on Data Mining (ICDM 2008) (pp. 263–272). DOI: 10.1109/ICDM.2008.22. Claim it supports: Formalizes the distinction between explicit and implicit feedback and the confidence-weighted approach used when only implicit signals exist. Wearly's deliberate choice to use only explicit feedback (the reject button, the wear-confirmation) is positioned against this alternative — cited so the choice is honest about what we are not doing. Relevant rule(s): wear-history-rules.md R6 (confirmation triggers a write); privacy-guidelines.md R1 (minimum-necessary access).

Verification status. verified citations.

3.9 Privacy and personal data¶

Definition. Legal, ethical, and design frameworks for handling personal data in a respectful, lawful, and minimum-necessary way.

Why it matters for Wearly. Calendar data, wardrobe data, body-shape declarations, modesty preferences, and wear history are all sensitive. The product brief commits to "no accounts, no third parties" in the prototype and to encrypted, opt-in, deletable data in production.

Source types to gather. - GDPR (EU Regulation 2016/679) is a publicly available legal framework. [to verify article numbers] - CCPA / CPRA (California). [to verify] - ACM / IEEE design principles for privacy-respecting AI systems. - Privacy-by-design framework (Ann Cavoukian) — well-known framework. [to verify exact title and year]

Current basis. Documented in book/10-privacy-security.md. The prototype follows minimum-necessary access, no third-party transmission (except the unauthenticated weather call), and no implicit data collection.

Verified citations.

Cited: Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 (General Data Protection Regulation). EUR-Lex CELEX:32016R0679. Official text: eur-lex.europa.eu/eli/reg/2016/679/oj. Specifically: Article 15 (right of access), Article 17 (right to erasure / "right to be forgotten"), Article 20 (right to data portability), Article 21 (right to object). Claim it supports: The four user-rights primitives Wearly's privacy stance commits to (view your data, delete your data, export your data, object to a use) map to named GDPR articles. Wearly's prototype does not implement a compliance pipeline; it implements the spirit of these articles by keeping data local-only, exportable as JSON, and trivially deletable. Relevant rule(s): privacy-guidelines.md R6 (no data retention beyond the session); book/10-privacy-security.md.

Cited: Cavoukian, A. (2009). Privacy by Design — The 7 Foundational Principles. Information & Privacy Commissioner of Ontario. Adopted as international standard by the 32^nd International Conference of Data Protection and Privacy Commissioners (Jerusalem, 2010). Public PDF: sfu.ca/~palys/Cavoukian-2011-PrivacyByDesign-7FoundationalPrinciples.pdf. Claim it supports: Seven principles including proactive not reactive, privacy as the default, full functionality (positive-sum), end-to-end security, and user-centric design. Wearly's "no accounts, no third parties, no implicit tracking" stance is a direct expression of privacy as the default and user-centric design. Relevant rule(s): privacy-guidelines.md R1–R6; book/10-privacy-security.md.

Verification status. verified citations for GDPR and Privacy-by-Design; [to verify] for CCPA/CPRA section references if/when the prototype ever transmits to US-jurisdiction servers (currently not applicable — single-user, local-only).

4. Body-shape framing — a deliberate note¶

The seed wardrobe (wardrobe.json) carries an owner.body_shape field with values like "hourglass". The Skill rule pack (skills/wearly-styling-agent/fit-silhouette-rules.md) references these labels in proportion-related guidance.

This is a user-declared proportion preference, not a body classification Wearly performs. Specifically:

The agent never analyzes user images to infer body shape. There is no image classifier, no body-detection model, no measurement extraction.
The shape label is only a proxy for a small set of proportion-related styling rules ("if hourglass → don't hide the waist when the user wants it highlighted").
The shape taxonomy itself (hourglass / pear / apple / inverted triangle / rectangle) is industry heuristic, not scientific taxonomy. It has no agreed academic standing. We use it because users recognize it; we don't claim it's a valid classification of bodies.
A user can leave body_shape blank, and no proportion-rule fires. Wearly works fine without it.

The right framing in any user-facing text is: "if you've shared a body shape preference" — never "based on your body type" (which implies classification) and never anything corrective.

The body-positive language contract (fit_tool.FORBIDDEN_TOKENS) enforces this at runtime. The static documentation check (tests/test_documentation_integrity.py) enforces it across this document, the Intelligent Book, and the Skill package.

5. Color framing — a deliberate note¶

The color_rules.json palettes (warm olive, cool fair, deep warm) are design-informed conventions, not deterministic claims:

"Best for warm olive" means "the warm/cool color-theory convention suggests these will read well; many designers agree." It does NOT mean "empirical research has shown these colors are objectively most flattering."
The +8 / +4 / −10 scoring formula is a UX choice: enough granularity to surface flags meaningfully, simple enough to fit on one card.
The honest claim in the reasoning trail is "this color tends to work well for warm-olive undertones" — calibrated language, not certainty.

6a. How reasoning lines cite rules (the citation chain)¶

Wearly's recommendation outputs are designed to be traceable end-to-end:

evidence category  ->  Skill rule pack  ->  rule ID  ->  reasoning line
   (this doc §3)         (skills/*.md)        (R<N>)      (result.reasoning)

The mechanism is a single registry, rule_refs.py, holding canonical slugs of the form <pack>#R<N> (e.g. occasion#R3, weather#R4, color#R2). Each slug names:

the Skill file it points at,
the R<N> heading inside that file,
a short body-positive one-liner summary.

styling_agent.py calls cite('occasion#R3') while building reasoning lines, which produces a compact trailing tag such as [occasion-rules#R3]. A UI layer can detect that tag and turn it into a deep link; a reviewer can grep for it; a documentation-integrity test asserts that every slug still resolves to a real rule heading. If a rule is renamed or renumbered without updating the registry, tests/test_rule_refs.py fails immediately — citations cannot silently drift.

This is the closest thing Wearly has to a single source of truth for "why this line is in the output." The Skill rule packs are that source; the registry is the index that keeps the runtime and the documentation aligned.

What the registry deliberately does NOT do:

It does not duplicate rule prose. The body of each rule lives in its *.md file and is not copy-pasted.
It does not fail recommendations on a missing slug. cite() returns "" for unknown slugs so a stale tag never crashes a result.
It does not claim every reasoning line carries a citation. Step 2 (load profile), Step 4 (filter announcement), and some fit-alignment notes are descriptive rather than rule-driven and stay unannotated. The test only asserts at least one citation per default occasion.

6. How Wearly's rule packs map to evidence categories¶

Skill rule pack	Primary category	Secondary categories
`occasion-rules.md`	Outfit compatibility (§3.2)	Fashion recommender systems (§3.1)
`weather-rules.md`	Outfit compatibility (§3.2)	—
`fit-silhouette-rules.md`	Body-shape-aware / fit-aware styling (§3.4)	Human-centered AI (§3.6)
`wardrobe-filtering-rules.md`	Wardrobe management (§3.7)	Personalization (§3.8)
`color-coordination-rules.md`	Color harmony (§3.3)	—
`wear-history-rules.md`	Wardrobe management (§3.7)	Personalization (§3.8)
`shopping-gap-rules.md`	Fashion recommender systems (§3.1)	Wardrobe management (§3.7)
`privacy-guidelines.md`	Privacy and personal data (§3.9)	Human-centered AI (§3.6)

Each skill rule file carries a Source basis footer pointing to this table.

7. Verification status summary¶

Category	Status	Verified citations	Pending
3.1 Fashion recommender systems	`verified citations`	1 (Ricci et al. 2022)	Fashion-XAI user study
3.2 Outfit compatibility	`verified citations`	2 (Han et al. 2017; Vasileva et al. 2018)	Deployed-product whitepaper
3.3 Color harmony	`verified citations`	3 (Itten 1961/1970; Munsell 1905; ISO 11664-4:2008)	Empirical skin-tone × colour study
3.4 Body-shape / fit-aware	`verified citations`	1 (Hokka 2024)	Direct critique of body-shape taxonomies
3.5 Explainable recommendation	`verified citations`	1 (Miller 2019)	Fashion-specific XAI user study
3.6 Human-centered AI	`verified citations`	2 (Amershi et al. 2019; Google PAIR Guidebook)	—
3.7 Wardrobe management	`general practice`	0	Sustainable-wardrobe research, industry whitepapers
3.8 Personalization / feedback	`verified citations`	1 (Hu, Koren & Volinsky 2008)	Explicit-feedback user study
3.9 Privacy and personal data	`verified citations`	2 (GDPR 2016/679; Cavoukian 2009 Privacy by Design)	CCPA/CPRA references (only if jurisdiction applies)

Ten verified citations across nine categories. Eight of the nine categories now carry at least one verified primary source; §3.7 (Wardrobe management / digital closets) remains general practice because no peer-reviewed source has yet been located for the wear-history rotation heuristics — see its Pending items.

Every cited work was located by direct search and its bibliographic details verified against the publisher or an indexing service (Crossref, ACM DL, IEEE Xplore, EUR-Lex, ISO). No citations are invented.

8. How to add a new citation¶

When a future contributor verifies a citation, the workflow is:

Identify the category above where the citation belongs.
Add the citation under that category's "Current basis" or as a new sub-section using this format:

> **Cited:** [Author(s), Year]. *Title*. Venue. DOI / URL.
> **Claim it supports:** "<one-sentence claim from Wearly's rules>"
> **Relevant rule(s):** [skill rule pack name(s) or book chapter(s)]

Update the verification-status table in §7 (change general practice → verified citations once at least one citation is present).
Remove the relevant item from "Pending items" in that section.
Update the skill rule's Source basis footer if the rule's basis has shifted from general practice to a verified citation.

Citations should be specific enough to verify (DOI, URL, or full bibliographic entry) and directly support the Wearly rule that references them. Citations that are general background context without a direct mapping should be filed in Wearly_References_and_Course_Context.md instead.

9. What this document is not¶

Not an academic literature review. A real lit review for Wearly is future production work.
Not a guarantee of correctness. The categories are the right framework; specific claims still need verification.
Not a substitute for the body-positive language contract. The contract is enforced in code (fit_tool) and in a documentation test; this document explains the reasoning, not the enforcement.

Companion to Wearly_Product_Brief.md §11 (Scientific and Expert Logic), Wearly_References_and_Course_Context.md §7–§8, and book/12-evidence-and-references.md.