Learning — how the engine improves

The engine gets better because every grade Shawn gives it changes what it does next. A grade is not a label filed away for later; it is an instruction the system acts on within seconds. This page covers what learning is in the system and where it takes effect. For the mechanics of how a grade is turned into a clean note, see How a grade becomes a note.

One learning surface

All learning from grades lands in one place: the entity notes in notes/ — one file per designer, colorway, and garment, each with a Guidance section and a Ban section. This is deliberate. With a single surface, there is exactly one place to look when you want to understand a behaviour, and exactly one place a grade can change.

The notes were reset to a blank slate so the engine could relearn cleanly from real grades rather than carry forward old noise. They fill back in, one clean construct at a time, as Shawn grades.

The other two layers of the brain — the curated brand law and the static vocabulary — do not learn from grades. Brand law is changed only by a person; the vocabulary is a fixed, on-brand word-bank. Keeping learning to a single surface is what keeps the brain coherent.

Bans require Shawn

The engine learns guidance on its own, but a ban (kill phrase) is only created when Shawn's grade note names what to avoid — the model never suggests one. A ban permanently fences off language across every future render, so it demands a human's explicit instruction; a wrong ban is far costlier than a wrong suggestion. An off-brand grade with no note records no ban. This is enforced at two layers — the grade analyser drops model-inferred bans, and the note writer drops any ban not authored by Shawn.

Where learnings take effect

Once a learning is in the notes, it shapes future work in several places:

1. Compose — the Guidance notes for the manifest's exact designer, colorway, and garment are read into the prompt that builds the next render.
2. Negative prompts — Ban triggers are added as short noun tokens that steer the image model away from the construct that failed.
3. Render-hint variants — banned fabric and detail language is skipped when the engine proposes how to render a garment.
4. The Lab — winning grades seed the experiments that clone and evolve a look; the notes scope what each experiment may and may not do.
5. Discovery — proven directions are reinforced and recent designs are fenced off, so the engine keeps exploring rather than repeating itself.

Garment-level learning

In ReviewHub, clicking a single segmented garment grades that garment on that model. The lesson lands scoped to the garment file, and the record knows which model slot it came from — so the system learns which piece of an outfit won, not merely that an image looked good.

Trust and provenance

Every note carries its source, and that source determines authority when the engine reads notes back. A learning Shawn states explicitly outranks one the system inferred on its own, and both are visible and editable in the Strategist's Brain tab. Because the notes are plain files under git, the full history of what was learned, when, and why is always available.