How a grade becomes a note¶
When Shawn grades a render, the system turns that judgement into a durable learning. This page explains exactly how — and, just as importantly, how the system guarantees that the learning is clean, correctly placed, and trustworthy.
The two-stage pipeline¶
A grade is processed in two stages, each handled by a focused model call.
Stage 1 — Vision. The grader looks at the actual rendered image alongside the grade and the prompt that produced it, and writes a forensic visual analysis: what rendered, what the prompt asked for, where they diverged, and the single phrase that most likely caused any drift. Because this stage sees the image, its observations are grounded in what is really on screen.
Stage 2 — Pattern. A second call reads that analysis and extracts the reusable lessons as structured notes — one atomic learning each, scoped to the designer, colorway, combination, or garment it concerns.
Splitting the work keeps each stage simple: the first describes, the second generalises.
What a clean note looks like¶
The brain holds two kinds of note, and only two:
- A Guidance note is one short, positive pattern — for example, "diamond quilting reads luxe." It is a single construct, never a list and never a sentence of explanation.
- A Ban note is a clean trigger phrase of a few words — for example, "visible silver zipper" — that is kept out of that entity's future prompts.
Every note concerns exactly one thing. If a render reveals three separate lessons, the system writes three separate notes.
The guarantees¶
The quality of the notes does not depend on the language model behaving perfectly. Every note passes through deterministic checks before it is written, and anything that fails is dropped rather than written badly. These checks are the reason the brain stays coherent over thousands of grades.
| Guarantee | What it prevents |
|---|---|
| Bans are Shawn's | A kill phrase is only created when Shawn's grade note names what to avoid. The model never suggests a ban on its own — an off-brand grade with no note teaches no ban. Enforced twice: the analyser drops model-inferred bans, and the note writer drops any ban not authored by Shawn. |
| Atomicity | A note must be a single construct. Multi-part text — joined by and, commas, or + — is rejected, so lessons never get crammed together. |
| Clean bans | A Ban must be a short trigger phrase. A negative note without one is dropped, so a sentence of description can never become a ban. |
| No self-negation | A note may not name its own designer, colorway, or garment. Banning the very thing a prompt specifies would cancel every render of it. |
| Grounding | A ban trigger must refer to something that actually appeared in the prompt or the visual analysis. The system cannot invent a ban for something that was never there. |
| No contradiction | A construct cannot be both guidance and a ban in the same file. If they collide, the ban wins — a proven failure outranks a praise. |
| Correct scope | A garment-level grade writes only garment notes; an image or model grade writes only designer, colorway, or combination notes. Lessons never land at the wrong level. |
| Safe content | Note text is checked for control characters and instruction-like phrasing, so a malformed or adversarial grade cannot smuggle anything into the prompts the engine builds. |
Where notes are written¶
Each note is written directly to its entity's file the moment the grade is processed, and the change is committed to git. There is no staging file and no nightly batch that has to run first — the learning is live immediately, scoped to the right designer, colorway, or garment.
Notes that the model infers without an explicit grade note from Shawn are marked as lower-trust, so a human's stated learning always outranks the system's own inference when the engine reads them back.
Genuinely universal lessons¶
Occasionally a lesson is truly independent of any one entity — for example, a
phrasing that the image model always misinterprets. Those lessons are routed,
conservatively, into the synthesised-craft files (craft_patterns.md or
render_quirks.md), which the engine reads on every compose. This is the only
automated path that touches a brand-law file, and it is restricted to those two
files alone; the immutable law can only be changed by a person.
What each grade does¶
| Grade | Effect |
|---|---|
| outlier / catalyst | Becomes guidance naming the winning construct; seeds the Lab to clone and evolve the look |
| on-brand | Usually records nothing — it confirms the system is already on target |
| off-brand / delete | Becomes a ban with the drift construct as its trigger — only if Shawn's note names what to avoid |
| ugly | Becomes a garment-scoped ban when Shawn's note names the broken phrase |
Why bans are held to a higher bar than guidance: a kill phrase permanently fences off language across every future render, while a piece of guidance only nudges. A wrong ban is far costlier than a wrong suggestion, so bans require a human's explicit instruction; guidance may still be inferred from a winning grade.