Learning — what a grade actually does¶
A grade is not a label that sits in a database; it is an instruction the system acts on immediately. Within seconds of a grade landing in ReviewHub, the following happens:
| Grade | Immediate effects |
|---|---|
| outlier | birth certificate (Lab clone-priority seed) + vocab burst (clone intent) + LLM analysis → notes |
| catalyst | birth certificate (evolve seed) + vocab burst (bridge intent) + notes |
| on-brand | birth certificate (validated-bounds seed) + notes |
| off-brand / ugly | kill phrases extracted → entity notes → injected into future negative prompts + sanitizer |
| delete | excluded from learning corpora |
Where learnings live¶
File-backed memfs (notes/): per-designer, per-colorway, per-garment, and
universal.md. Each note carries type (recipe = guidance the LLM reads,
kill_phrase = hard ban) and provenance (Shawn vs self-learned). Notes are
git-versioned on the volume for audit history.
Where learnings act¶
- Compose system prompt — scoped notes for the manifest's exact entities
- Negative prompts — kill-phrase triggers as short noun tokens (filtered so they never negate the look's own entities)
- Render-hint variants — banned fabric/detail variants skipped
- Lab — seeds, craft blocks, per-seed scoped notes, fuse-guards
- Discovery — proven directions + anti-repetition fence
- Sleeptime synthesis — periodic clustering of grade evidence into universal craft rules (n≥2 promotion threshold)
Garment-level grades¶
Clicking a segmented garment in ReviewHub grades THAT garment on THAT model:
the learning lands scoped to the garment (J26.md style) and the birth
certificate carries the model slot — the Lab knows which outfit won, not just
which image.