04 · Styling knowledge base

The app applies these rules at runtime; this chapter is the reference layer that explains each rule pack and the data it reads.

Entry point to Part II. This chapter is the structural map of what Wearly knows. Each rule system below is taught in its own Part-II chapter, learner-first, with the evidence behind it:

• Color → §S1 · Color and skin tone
• Silhouette / fit → §S2 · Silhouette and fit
• Occasion → §S3 · Occasion and context
• Weather / layering → §S4 · Weather and layering
• Construction → §S5 · Wardrobe construction
• Rotation → §S6 · Rotation and freshness
• Gaps → §S7 · Honest gaps
• Body-positive framing → §S8 · Body-positive framing as a discipline

Every Wearly decision traces to an inspectable knowledge base — three rule systems, eight rule packs, and a citation registry. This chapter documents what's in it and why.

Three rule systems

1. Occasion → tag map

In styling_agent.py:

OCCASION_TAG_MAP = {
  "work", "business", "meeting" → "work",
  "gym", "workout", "yoga", "running" → "gym",
  "dinner", "restaurant", "date" → "dinner",
  "formal_event", "gala", "formal" → "formal",
  "casual", "weekend", "brunch", "everyday" → "casual",
}

A free-text occasion like "client dinner" is matched against this map. The agent always lands on one of five canonical tags. This means the wardrobe filtering, the dress-vs-separates decision, and the required-pieces set are all driven by a small, well-known vocabulary.

2. Required pieces per occasion

REQUIRED_PIECES = {
  "work":    ["top", "bottom"],
  "dinner":  ["top", "bottom"],   # unless a dress is preferred
  "gym":     ["activewear", "activewear"],
  "formal":  ["dress"],
  "casual":  ["top", "bottom"],
}

Step 6 (gap check) compares the built outfit against this list. A missing required piece becomes a wardrobe gap.

3. Color rules per skin tone

color_rules.json defines best_colors, good_colors, avoid_colors, metal_preference, and a notes line for each of three skin tones today:

• warm olive — terracotta, camel, burgundy, deep teal, mustard, gold
• cool fair — navy, icy blue, lavender, silver, emerald, raspberry
• deep warm — cobalt, emerald, bright red, gold, orange, purple

The scoring formula is intentionally transparent — base 60, +8 for best, +4 for good, −10 for avoid, clamped to 0–100. Every contribution is visible in the reasoning trail; a reviewer can verify each item's score by inspection — a property a deep-learned color model would not have.

Why the reasoning core is inspectable rules

The course principle "structure beats volume" applies directly. A learned compatibility model would:

• need large labeled datasets to train,
• be opaque ("the network said so"),
• be impossible to trace from a reasoning line back to a citable rule,
• prevent the body-positive language contract from being enforced at the prose layer.

A structured rule system, by contrast:

• is fully visible across eight rule packs, the citation registry (rule_refs.py), and the Wearly Knowledge Graph — every decision traces to a named rule,
• can be edited at runtime — rule changes take effect on the next recommendation, with the new reasoning trail showing exactly what changed,
• supports an audit trail without extra plumbing,
• is the strongest foundation for adding learned layers later — the rule outputs become labeled training data for any future model, and the safety + scoring substrate stays in place underneath.

Next rule layers on the roadmap

• Body shape → silhouette (Phase 3, fit profile).
• Modesty preferences (Phase 3).
• Comfort and fabric (Phase 4).
• Wear history (Phase 4 — a piece of this is shipped today).
• Cross-item color harmony (currently per-item only).
• Brand affinity / aesthetic clusters (Phase 5).

Each is a future rule layer, not a future ML feature. The structure stays inspectable as it grows.

What informs these rules

Each rule system is evidence-informed, design-informed, and user-preference-driven — not claimed to be exact or objectively correct. The canonical sourcing lives in docs/evidence-and-references.md. Quick map:

Rule system	Primary evidence category	Current basis
Occasion → tag map	Outfit compatibility (§3.2)	Industry styling heuristics — five canonical occasion types are widely recognized in fashion-product UX.
Required-pieces map	Outfit compatibility (§3.2)	Industry styling heuristics.
Color palettes	Color harmony (§3.3)	Design-informed conventions (warm / cool color-theory). The "Best for warm olive" claim is "these tend to read well under warm/cool conventions" — not an empirical claim.
Color scoring formula (+8 / +4 / −10)	Color harmony (§3.3)	UX choice for readability — not derived from a specific empirical model.

The honesty contract from chapter 12: Wearly says "tends to work well," not "is correct."

Where to verify

• The maps: styling_agent.py (OCCASION_TAG_MAP, REQUIRED_PIECES).
• The palettes: color_rules.json.
• The skin-tone scoring: color_tool.py · score_outfit_colors().
• The canonical sourcing: docs/evidence-and-references.md.

---

Key terms

• Required pieces — the minimum type-set an occasion needs (e.g., work = top + bottom) (glossary).
• Color tier — the +8 / +4 / −10 scoring scale for color against skin-tone palette (glossary).
• Skin-tone palette — one of three named palettes the color formula reads from (glossary).
• Rule citation (R) — a numbered heading inside a Skill rule file (glossary).

Self-check

1. (Remember) Name the three skin-tone palettes Wearly ships today.
2. (Understand) Why is Wearly's reasoning core structured as inspectable rules, and what does that enable that a learned model could not?
3. (Apply) Look up occasion-rules.md R2 in the skill pack. Which rule heading does it resolve to, and which chapter cites it?

Definitions live in the Glossary. Self-check questions follow Bloom's taxonomy progression (Remember → Understand → Apply → Analyze).