Glossary

Look — one generated image: three models, three complete outfits, one designer's aesthetic, one colorway.

Manifest — the dealt recipe for a look before any text is written: the designer, colorway, and the three model outfits chosen from the roster.

Roster — the curated pools the engine draws from: designers, colorways, garment bodies, hair styles, and outfit templates. Managed in Strategist → Control.

Outfit template — a pattern for one model's outfit: which garment category is the hero (jacket, dress, top, vest) and which slots get filled around it.

Hero garment — the featured garment of a model's outfit. Its silhouette is always spliced verbatim into the prompt; innovation happens around it.

Scene template — the locked stage: seamless white backdrop, three models evenly spaced, ultrawide 16:9, full-length framing. Identical on every look by design; a validator flags any drift.

Innovation axes — the eight directions the engine invents along: collar architecture, closure systems, panel mapping, hem geometry, layering logic, texture pairing, hardware language, surface treatment. Two rotate per look.

Anti-repetition fence — construction phrases mined from recent looks that the next look is told not to reuse. Keeps "novel" from converging on yesterday's phrasing.

Grade — Shawn's verdict on a look or a single garment: outlier (winner), catalyst (spark worth evolving), on-brand (meets the bar), off-brand (violates the brand), ugly (unusable render), delete.

Birth certificate — the complete DNA record created when a look grades outlier / catalyst / on-brand: prompt, garments, designer×colorway, graded model slot, notes. The Lab's seeds.

Seed — a birth certificate waiting in the Lab's queue.

Kill phrase — a short banned phrase Shawn calls out in a grade note. Lands in the negative prompt of future renders. Created ONLY from Shawn's note — the model never suggests one; an off-brand grade with no note teaches no ban. Always a clean trigger of a few words; a sentence can never become a kill phrase.

Note — a single learned construct scoped to a designer, colorway, combination, or garment. Lives in one of two sections: Guidance (a positive pattern to follow) or Ban (a kill phrase to keep out). Each note concerns exactly one thing.

The brain — the three layers of Markdown files the engine reads: the curated brand law (system/), the learned notes (notes/), and the static vocabulary (vocab/). Browsable and editable in the Strategist's Brain tab.

Brand law — the system blocks that define what Anatomie is and how the engine must behave. Mostly immutable and changed only by a person; two files (craft_patterns, render_quirks) are a learned craft layer the engine may append to under audit.

Grounding — the rule that a learned ban must refer to something actually present in the prompt or the render's visual analysis, so the system cannot invent a ban for something that was never there.

Contradiction guard — the rule that a construct cannot be both guidance and a ban in the same file. If they collide, the ban wins.

The Lab — the evolution arm: clones proven recipes, blends winning garment bodies into new hybrids, and runs frontier experiments. Has its own daily quota, separate from roster generation.

Evolution chain — seed → the experiments it spawned → the grades the descendants earned. Visible on the Lab tab.

memfs — the file-backed knowledge store: markdown files for pools, notes, vocabulary, and brand rules. Git-versioned on a persistent volume. The body of a file is the memory; the engine reads these files when composing.

Negative prompt / Avoid line — the exclusion list a render receives: base brand bans + the look's specific exclusions + learned kill phrases, as short noun phrases.

Segmentation / hitmap — per-garment masks over a rendered image so ReviewHub can attribute a click (and a grade) to one garment on one model.

Pipeline events — the observability table: every stage of the pipeline records ok/fail with the reason. Powers the Status panel and health alerts.