Research AI answers a different question than Google:

Should I spend time on this topic tonight?

That answer — Learn Now, Learn Later, Skip, or Delegate — must not drift between layers.

The rule

research-ai-worker     → run_pipeline(), write v2:research:{hash}
research-ai-back-end   → cache_get, enqueue, rate limits
research.xingai.app    → render cached fields

We copied the lesson from Invest AI (ADR-012): one decision engine, one cache row.

What stays in the worker

• Verdict and one-sentence summary
• Sub-scores and Learning ROI (deterministic formula — see sibling post)
• 30-minute route, takeaways, ranked sources
• Discussions + knowledge graph enrichments
• Trending topic warm list

What FastAPI may do

Read SQLite KV, validate JSON, attach freshness metadata, enforce 3 live runs/day for anonymous users, sync saved library rows.

It does not call OpenAI to “fix” a cache miss on the request thread.

Why it matters for learning products

Without the boundary, a compare page could rank topic A above topic B using frontend math while the result page says the opposite. Share links would lie.

Users are allocating attention, not money — but the trust problem is the same.

Further reading: xingai-research-ai/docs/adr/002-decision-cache-boundary.md