FAQ

I graded something — when does it matter? Within seconds. Winners join the Lab's seed queue immediately; off-brand grades write kill phrases that fence the next compose. The next batch already reflects your grades.

Why does every image look staged the same way? By design. The scene template is locked so that grades measure the clothes, not the photography. A validator flags any prompt that drifts from the stage.

Where does genuinely new design come from? Two places. Roster looks innovate construction language (collars, closures, panels, surfaces) on proven garment bodies — that's the innovation engine. New garment bodies are born in the Lab, by blending two proven winners into a hybrid.

Can the engine repeat itself? It actively resists it: outfit-hash dedup against the last 200 looks, daily rotation of stale designers/colorways, and the anti-repetition fence on construction language. Repetition is reserved for the Lab's clone lane, where re-testing a winner is the point.

What happens if a model or API starts failing? Every stage records failures with reasons. If any stage fails ten times in a day, the health badge goes red and Status names the stage and the why. Nothing fails silently — that's a lesson the system learned the hard way.

Does anything live only inside an AI's memory? No. Every learning is a readable file (the memfs) or a queryable row (Postgres). The Learning tab browses the same files the engine reads.

What do I do before turning on daily automation? Confirm the persistent memfs volume exists on the deployment (see Deploy & verify), then flip "Runs automatically" in Control. The countdown next to the schedule confirms it's armed.

Why did a batch cost less / stop early? Two budget stops: the per-run render budget and the daily budget, both in Control. The engine stops cleanly when either is hit.