The Lab¶
If roster generation is the system's curiosity, the Lab is its memory of what worked. The roster explores broadly; the Lab takes the looks Shawn graded as winners and squeezes them — re-testing proven recipes in new contexts, fusing winning garments into new hybrids, and pushing careful experiments past the edge of the catalog.
Seeds¶
Every outlier / catalyst / on-brand grade creates a birth certificate: the look's full DNA (prompt, garments, designer×colorway, the graded model slot, grader notes). Re-grading the same render supersedes its seed (no duplicates; upgrades re-fire learning).
Experiment types¶
| Type | What it does | Heritage |
|---|---|---|
| clone / autopilot (APL) | re-tests a proven recipe — verbatim prose with a new designer/colorway, or a thematic focus mined from winners | verbatim — repetition is the job |
| mutate / re_pair | one variable swapped (designer, colorway, hero transplant) | preserved |
| blend | two proven garment BODIES fused into a new synthetic piece | new body born |
| explore / frontier (EXP) | pushes construction language past the catalog along the innovation axes | frontier — gets the full invention push |
The Lab agent receives the craft blocks, per-seed scoped notes, MAP-Elites niche map, and its own per-lane hit rates (which experiment lanes actually produce outliers) — second-layer learning that shifts its allocation.
Genealogy¶
Every experiment links to its parent seed (parent_birth_cert_id, mined from
the proposal when the agent names seeds in free text). Strategist → Lab shows
the chains: seed → experiments → the grades the descendants earned. When a
descendant grades outlier, it becomes a seed itself.
Quota¶
lab_batch experiments/day (Lab page; default 12; 0 parks the lab). Separate
from the roster batch. "Run lab now" fires a cycle on demand.
Guards¶
Lab output passes the same validators as roster prompts (learned bans, scene invariants). Spliced vocabulary carries grade attribution — fragments that grade poorly get auto-retired.