When a kid taps "ask Xplorer" or when our editorial team queues a lesson, a pipeline kicks in. Here's what happens in the 60-90 seconds before the lesson appears in-app.
Step 1: Queue (0-2s)
The question is enqueued with metadata: target age band, requested world, source (child-asked vs editorial). Critical: no identifying child data leaves the app.
Step 2: Framing model call (5-15s)
A first model pass reframes the question into the lesson template: title, hook, three sub-lessons, question per sub-lesson. The prompt includes the age-band voice guide.
Step 3: Content safety pass (3-8s)
Classifier sweep for medical/political/violence/romance content. Anything over threshold goes to a human review queue. Nothing surfaces to a child pre-review.
Step 4: Age-band variants (10-20s)
The base lesson is auto-specialised into easy / medium / hard difficulty within the target age band. Each sub-lesson gets three question variants.
Step 5: Illustration (15-30s)
Feature image + optional sub-lesson illustrations are generated. Safety pass on images too.
Step 6: Publish (1-3s)
Lesson is stored, indexed into search, associated with its world and age band, and surfaced. Edit history is retained for audit.
The actual cost
A full lesson costs us about $0.12 to produce. The vast majority of that is the illustration step. The framing + safety steps run on cheap, fast models because the lesson structure is rigid.