The brain — how the engine knows what it knows

Everything the engine understands about the Anatomie brand lives in a single, inspectable place: a set of Markdown files called the memfs (memory filesystem). There is no hidden model state and no opaque database of embeddings. If you want to know why the engine made a decision, you can open the exact file it read.

The brain is organised into three layers, each with a clearly different job and a clearly different owner.

The three layers

1. Brand law — the system blocks

The files in system/ define what Anatomie is and how the engine must behave. They are the constitution, and they are injected into every compose:

File	What it governs
brand_dna.md	Identity, approved fabrics, the colour family and colour-lock
brand_guardrails.md	Every hard ban — killed fabrics, silhouettes, and styles
construction_rules.md	How garments are built (quilting, hybrids, pant rules)
prompt_base.md	The compose contract — output schema, scene lock, structure
craft_patterns.md	Proven recipes and winning combinations (a learned layer)
render_quirks.md	Phrasings that misfire in the image model, regardless of brand

The first four are immutable law. They are curated by hand and changed only by a person, through the Strategist's Brain tab. The engine cannot rewrite them.

The last two — craft_patterns and render_quirks — are the synthesised craft layer. They hold knowledge the system accumulates from experience, and the engine is allowed to append to them under strict, audited conditions (see How a grade becomes a note).

2. Learning — the entity notes

The files in notes/ are the engine's living memory of each individual designer, colorway, and garment. There is one file per entity (notes/designer/D04.md, notes/colorway/C09.md, notes/garment/J26.md), and each file has exactly two sections:

• Guidance — short, positive patterns to follow for that entity.
• Ban — clean trigger phrases to keep out of that entity's renders.

This is the single learning surface. When Shawn grades a render, the lesson is written here, scoped to the precise entity it concerns. Nothing else in the system learns from grades — which is what keeps the brain coherent.

The notes were deliberately reset to a blank slate so the system could relearn cleanly from real grades, rather than carry forward years of accumulated noise.

3. Vocabulary — the curated word-bank

The files in vocab/ are a static, hand-curated bank of on-brand fabric and construction phrases. The engine draws from them to vary the language of each render so that no two looks read identically.

The vocabulary does not learn from grades and has no performance counters. It is simply a clean set of words that are always safe to use. (Earlier versions ran a self-evolving vocabulary; it was retired because a self-mutating word pool drifts off-brand over time, and the entity notes already capture what to avoid.)

Why three layers

Each layer answers a different question, and keeping them separate is what makes the brain trustworthy:

• Brand law answers "what is allowed?" — and only a human changes it.
• Notes answer "what have we learned about this entity?" — grades write here.
• Vocabulary answers "how do we say it?" — curated, never auto-changed.

A single learning surface means there is exactly one place to look when a behaviour surprises you, and exactly one place a grade can change.

Browsing and editing the brain

The Strategist's Brain tab is a complete window onto the memfs. From it you can:

• Browse every file, grouped by layer — Memories, Brand law, Vocabulary, Skills.
• Open any file and read precisely what the engine reads.
• Edit any Markdown file; every save is committed to git for a full audit trail.
• Chat with an assistant that is grounded in the live files and can explain why the engine behaves the way it does, quoting the relevant lines.

Because the brain is just files, the Brain tab is the natural place to teach the engine a new rule: edit the relevant brand-law file, save, and the next compose reads the change — no redeploy required.