Google answers what. Research AI answers whether tonight:

Should I spend an hour on MCP, RAG, or neither?

One screen, one outcome

Each query returns a decision board:

• Verdict — Learn Now · Learn Later · Skip · Delegate
• Learning ROI (0–100)
• 30-minute route + ranked sources
• Optional graph + hot discussions

No chat thread. No essay by default. If users want depth, they open sources — not another prompt box.

Why we rejected chat-first

Chat hides the verdict below scroll. It invites follow-ups before the user commits attention. Our upgrade rule says inherit the main flow — for Research AI, the main flow is decide, not converse.

Backend follows the same story

Worker-owned inference, read-only API (ADR-002). The UI renders cached fields; it does not re-rank or re-decide.

Optional paths stay optional

Compare two topics. Portfolio by ROI. History in localStorage. Topic SEO pages for shareable briefs. None of them replace the home → verdict path.

Further reading: xingai-research-ai/docs/adr/001-learning-decision-system.md